JP7743053B2 - Graph generation device, state prediction device, graph generation method, state prediction method, and program. - Google Patents

Description

The present invention relates to a graph generation device, a state prediction device, a graph generation method, a state prediction method, and a program.

Several technologies related to condition prediction have been proposed. For example, Patent Document 1 describes acquiring dietary information and health condition information for multiple users, and predicting the future health condition of a first user based on the health condition information of a second user who is older than the first user, if the first user's dietary habits and the second user's dietary habits have a predetermined relationship.

Japanese Patent Application Laid-Open No. 2020-177469

If the predicted state change can be illustrated, the state change can be visually grasped. In this case, it is preferable that the time required for the state change is also shown on the diagram.

One example of the objective of the present invention is to provide a graph generation device, a state prediction device, a graph generation method, a state prediction method, and a program that can illustrate predictions of state changes and show the relative length of time required for a state change.

According to a first aspect of the present invention, a graph generation device includes: a data acquisition unit that acquires a time series of data related to a target for prediction; a graph generation unit that converts each of the data included in the time series into a feature using a feature extraction method obtained by learning using an evaluation index in which the closer the distance in a feature space between two feature values obtained by performing feature extraction on each of two pieces of data included in the same time series is to a value obtained by multiplying a predetermined proportionality coefficient by the time interval in the time series of the two pieces of data , the higher the evaluation, and generates a graph in which the feature space is a coordinate space; and a state information addition unit that adds, to a portion of the coordinate space, information indicating the state of the target for prediction at a time in the time series of the data converted into coordinate values included in that portion.
According to a second aspect of the present invention, a graph generation device includes a data acquisition unit that acquires a first graph in which a time series of data related to a prediction target is shown in a first coordinate space; a graph generation unit that performs coordinate transformation from the first coordinates to second coordinates using a distance index that is a weighted sum of the distance in the first coordinate space between two pieces of data included in the time series shown in the first coordinate space and the time interval in the time series of the two pieces of data, such that the smaller the value of the distance index, the closer the distance between the two pieces of data in the second coordinate space; and a state information addition unit that adds, to a portion of the coordinate space, information indicating the state of the prediction target at a time in the time series of the data converted to coordinate values included in that portion.

According to a third aspect of the present invention, a state prediction device includes a prediction unit that performs state prediction on a time series of data related to a prediction target using a graph in which the feature space is a coordinate space obtained by converting each of the data included in the time series into the feature using a feature extraction method obtained by learning using an evaluation index in which an evaluation is higher the closer the distance in feature space between two feature values obtained by extracting feature values from each of two data included in the same time series of the data related to the prediction target is to a value obtained by multiplying the time interval in the time series of the two data by a predetermined proportionality coefficient.

According to a fourth aspect of the present invention, a graph generation method includes: a computer acquiring a time series of data related to a target to be predicted; converting each of the data included in the time series into a feature using a feature extraction method obtained by learning using an evaluation index in which the closer the distance in a feature space between two feature values obtained by performing feature extraction on each of two pieces of data included in the same time series is to a value obtained by multiplying a predetermined proportionality coefficient by the time interval in the time series of the two pieces of data, the higher the evaluation ; and generating a graph in which the feature space is a coordinate space; and adding, to a portion of the coordinate space, information indicating the state of the target to be predicted at a time in the time series of the data converted into coordinate values included in that portion.

According to a fifth aspect of the present invention, a state prediction method includes: a computer converting each of the data included in the time series into features using a feature extraction method obtained by learning using an evaluation index in which an evaluation is higher the closer the distance in a feature space between two feature values obtained by performing feature extraction on each of two data included in the same time series of data related to a prediction target is to a value obtained by multiplying a time interval in the time series of the two data by a predetermined proportionality coefficient, and predicting a state of the prediction target using a graph in which the feature space is a coordinate space .

According to a sixth aspect of the present invention, a program causes a computer to execute the following steps: acquire a time series of data related to a target to be predicted; convert each of the data included in the time series into a feature using a feature extraction method obtained by learning using an evaluation index in which the closer the distance in a feature space between two feature values obtained by performing feature extraction on each of two pieces of data included in the same time series is to a value obtained by multiplying a predetermined proportionality coefficient by the time interval in the time series of the two pieces of data , the higher the evaluation; and generate a graph in which the feature space is a coordinate space; and add, to a portion of the coordinate space, information indicating the state of the target to be predicted at a time in the time series of the data converted into coordinate values included in that portion.

According to a seventh aspect of the present invention, there is provided a program for causing a computer to execute a state prediction of a prediction target using a graph having the feature space as a coordinate space, the graph being obtained by converting each of the data included in the time series into the feature using a feature extraction method obtained by learning using an evaluation index in which an evaluation is higher the closer the distance in feature space between two feature values obtained by performing feature extraction on each of two data included in the same time series of data related to the prediction target is to a value obtained by multiplying the time interval in the time series of the two data by a predetermined proportionality coefficient.

The graph generation device, state prediction device, graph generation method, state prediction method, and program described above make it possible to visualize predicted state changes and show the relative length of time required for a state change.

FIG. 1 is a diagram illustrating an example of the configuration of a graph generation device according to a first embodiment. FIG. 2 is a diagram illustrating an example of the configuration of a graph generation unit according to the first embodiment. FIG. 10 is a diagram illustrating an example of a graph in which the time represented by the distance in the feature space differs depending on the position in the feature space. FIG. 2 is a diagram illustrating a first example of the configuration of a feature extraction unit according to the first embodiment. FIG. 10 is a diagram illustrating a second example of the configuration of the feature extraction unit according to the first embodiment. FIG. 2 is a diagram showing an example of a graph output by the graph generation device according to the first embodiment. FIG. 4 is a diagram showing a first example of a state indicated by a state change graph according to the first embodiment. FIG. 10 is a diagram showing a second example of a state indicated by a state change graph according to the first embodiment. FIG. 10 is a diagram showing a third example of a state indicated by a state change graph according to the first embodiment. FIG. 10 is a diagram showing a fourth example of a state indicated by a state change graph according to the first embodiment. FIG. 10 is a diagram showing a fifth example of a state indicated by a state change graph according to the first embodiment. FIG. 10 is a diagram showing an example of a state change graph in which the display area of an edge is determined depending on the direction of the edge in the first embodiment. FIG. 10 is a diagram showing an example in which a state change graph is displayed in a three-dimensional feature amount space in the first embodiment. 10 is a flowchart illustrating an example of a processing procedure in which the graph generation device according to the first embodiment performs learning of a graph generation unit 192. FIG. 10 is a diagram illustrating an example of the configuration of a graph generation device according to a second embodiment. FIG. 10 is a diagram illustrating an example of the configuration of a feature amount extraction unit according to the second embodiment. FIG. 10 is a diagram illustrating an example of the configuration of a state prediction system according to a third embodiment. FIG. 10 is a diagram illustrating an example of the configuration of a state prediction device according to a third embodiment. FIG. 11 is a diagram illustrating an example of nodes in a state change graph that are referenced by a state prediction device in the third embodiment. 10 is a flowchart illustrating an example of a process in which a state prediction device according to a third embodiment predicts a state of a prediction target. FIG. 10 is a diagram illustrating an example of the configuration of a graph generation device according to a fourth embodiment. FIG. 13 is a diagram illustrating an example of a graph before conversion in the fourth embodiment. FIG. 13 is a diagram illustrating an example of a graph after conversion in the fourth embodiment. FIG. 11 is a diagram showing a first example in which the graph generation device according to the fourth embodiment generates a graph from a power flow diagram of a state change speed vector. FIG. 11 is a diagram showing a second example in which the graph generation device according to the fourth embodiment generates a graph from a power flow diagram of a state change speed vector. FIG. 1 is a schematic block diagram illustrating an example configuration of a computer according to at least one embodiment.

The following describes embodiments of the present invention, but the following embodiments do not limit the scope of the invention as claimed. Furthermore, not all of the combinations of features described in the embodiments are necessarily essential to the solution of the invention.

First Embodiment
Fig. 1 is a diagram showing an example of the configuration of a graph generation device according to the first embodiment. In the configuration shown in Fig. 1, the graph generation device 100 includes a communication unit 110, a display unit 120, an operation input unit 130, a storage unit 180, and a control unit 190. The control unit 190 includes a data acquisition unit 191, a graph generation unit 192, a state information addition unit 193, and a learning control unit 194.

The graph generation device 100 generates a graph showing changes in the state of a graph generation target. The graph generation device 100 may be configured using a computer such as a workstation or personal computer (PC).
Specifically, the graph generation device 100 acquires a time series of m-dimensional data (time-series data) related to the state of the graph generation target, extracts n-dimensional features for each piece of data (each piece of data at each time in the time series of data), and generates a time series of the n-dimensional features. The graph generation device 100 represents each feature as a node in a feature space, and generates a graph in which edges represent the chronological relationship between the features in the time series.

The feature amounts referred to here are data converted from the original data. They are called "features" because the state can be determined based on the position where the feature amounts are plotted in the feature amount space.
Here, m and n are positive integer constants. Typically, m>n, but m≦n may also be satisfied.
The variables that represent the values of each dimension of a feature are also called latent variables. The feature space is an n-axis coordinate space with n latent variables. When n = 2, the graph generated by the graph generation device 100 is a planar (two-dimensional) graph. When n = 3, the graph generated by the graph generation device 100 is a three-dimensional (three-dimensional) graph.

The graph generation target here is not limited to a specific one, but can be various ones whose state can be determined based on the obtained data. Furthermore, the graph generation target here may be a single one, or a set, such as a set of the same type of thing. In particular, the graph generation target may be the same as the prediction target. Alternatively, the graph generation target may be another one or a set that is thought to undergo a state change similar to the prediction target.
The graph generation target here is a target whose state is referred to in order to generate a graph by the graph generation device 100. The prediction target here is a target whose state is predicted using the graph generated by the graph generation device 100.

For example, the subject of graph generation may be a person, and the graph generation device 100 may generate a graph for predicting the person's health state, various disease states, and other conditions. Furthermore, for example, the subject of graph generation may be measurement data of microRNA (miRNA), and the graph generation device 100 may generate a graph for predicting cancer.

When the graph generation target is a person, the graph generation device 100 may acquire time-series data for each of multiple people and generate a graph, which is expected to enable the graph generation device 100 to generate a graph for a wider area in the feature space, which is the coordinate space of the graph display target.
Here, although changes in conditions such as illness vary from person to person, there are certain trends, and it is expected that the graph generation device 100 will be able to generate graphs that are useful for predicting the condition of an illness, etc. The graph generation device 100 may generate a graph using time-series data of a group of people whose changes in condition are considered to be relatively similar, such as data categorized by age.

Alternatively, the graph generation target may be a specific machine, and the graph generation device 100 may generate a graph for predicting the normal state and various abnormal states of that machine. If the graph generation target is a specific machine and there are multiple machines of the same model used in similar environments, the graph generation device 100 may acquire time-series data from each of these multiple machines and generate a graph. This is expected to enable the graph generation device 100 to generate graphs for a wider area within the feature space, which is the coordinate space of the graph display target.

Alternatively, the graph generation target may be a specific mechanical system such as a factory production line, and the graph generation device 100 may generate a graph for predicting the normal state and various abnormal states of that mechanical system. Alternatively, the graph generation target may be a specific plant such as a chemical plant or power plant, and the graph generation device 100 may generate a graph for predicting the normal state and various abnormal states of that plant.

Alternatively, the graph generation target may be market conditions (trading conditions in a market such as the stock market), and the graph generation device 100 may generate a graph for predicting conditions such as an increase or decrease in trading prices.
In the following, an example will be described in which the prediction target is a person, and the graph generation device 100 generates a graph for predicting the person's health state and various disease states.
A graph representing a change in state generated by the graph generation device 100 is also called a state change graph.

The communication unit 110 communicates with other devices. For example, the communication unit 110 may communicate with a database that stores data related to the target for graph generation, and receive a time series of data related to the state of the target for graph generation.

The display unit 120 has a display screen such as a liquid crystal panel or an LED (Light Emitting Diode) panel, and displays various images. For example, the display unit 120 may display the state change graph generated by the graph generation device 100.
The operation input unit 130 includes input devices such as a keyboard and a mouse, and receives user operations. For example, the operation input unit 130 may receive a user operation to instruct the generation of a state change graph and a user operation to indicate a data acquisition source.

The storage unit 180 stores various types of data. For example, the storage unit 180 may store a time series of m-dimensional data related to the state of the graph generation target, and a time series of n-dimensional features. The storage unit 180 may also store a learning model for learning a feature extraction method. The storage unit 180 is configured using a storage device included in the graph generation device 100.

The control unit 190 controls each unit of the graph generation device 100 to perform various processes. The functions of the control unit 190 are performed, for example, by the CPU (Central Processing Unit) of the graph generation device 100 reading and executing a program from the storage unit 180.

The data acquisition unit 191 acquires various data used by the graph generation device 100 to generate a state change graph. In particular, the data acquisition unit 191 acquires a time series of data related to the prediction target. For example, as described above, when the communication unit 110 receives a time series of data related to the state of the graph generation target from another device, the data acquisition unit 191 may read data related to the state of the graph generation target from the data received by the communication unit 110.

The graph generation unit 192 generates the state change graph described above for the graph generation device 100. In doing so, the graph generation unit 192 generates a state change graph in which the length of an edge connecting two nodes with a straight line is proportional to the length of time in the time series data. In other words, the graph generation unit 192 converts each piece of data included in the time series of data related to the prediction target into coordinate values in a coordinate space set so that the length of time in the time series is proportional to the distance in the coordinate space, and generates a graph in the coordinate space.

As described above, the data conversion performed by the graph generation unit 192 can be considered feature extraction, and the coordinate space of the state change graph can be considered feature space. In the first embodiment, the graph generation unit 192 learns a feature extraction method. This learning refers to machine learning, and the graph generation unit 192 adjusts the parameter values of the learning model for the feature extraction process based on training data.

Fig. 2 is a diagram showing an example of the configuration of the graph generation unit 192. In the configuration shown in Fig. 2, the graph generation unit 192 includes a feature extraction unit 210 and a graph formation unit 220. Also shown in Fig. 2 is a learning control unit 194.
The feature extraction unit 210 performs the data conversion described above for the graph generation unit 192. The time series data input to the feature extraction unit 210 can be expressed as a vector as shown in equation (1).

fts represents the time series data input to the feature extraction unit 210. The time series data input to the feature extraction unit 210 is also referred to as input time series data. i is an identification number that identifies the time series data. Here, it is assumed that k pieces of time series data (k is a positive integer) are input to the feature extraction unit, where 1≦i≦k.
The graph generation device 100 may use multiple pieces of time series data obtained for different graph generation targets, or may use multiple pieces of time series data obtained at different times for the same graph generation target.

An example of a case in which the graph generation device 100 uses multiple pieces of time-series data obtained for different graph generation targets is when the graph generation device 100 generates a graph for predicting a person's health state, various disease states, etc., based on data for multiple people.An example of a case in which the graph generation device 100 uses multiple pieces of time-series data obtained at different times for the same graph generation target is when the graph generation device 100 generates a state change graph for predicting the normal state, various abnormal states, etc., of a single plant, based on data for that plant.
In the following, an example will be described in which the graph generation device 100 uses time-series data of k people. i is also used as an identification number for identifying the graph generation target (here, person), and it is assumed that the i-th time-series data is obtained for the i-th graph generation target.

The time series data used by the graph generation device 100 during learning and graph generation may be the same or different. Graph generation is also referred to as operation of the graph generation device 100. The following describes an example in which the graph generation device 100 uses the same time series data during learning and graph generation.

In the learning data, data for each time included in the time series data may be linked to information indicating the state of the subject for which a graph is to be generated at that time. The information indicating the state of the subject for which a graph is to be generated may be information indicating the name of a class determined using a state change graph, such as "healthy," "pre-illness," or the names of various diseases.

The data for each time included in the input time series data fts _i is also referred to as input data for each time. The input data at time t _i,j is also written as f _i (t _i,j ). _{f i} (t _i,j ) can be expressed as a vector as shown in Equation (2).

xmax is a positive integer constant representing the number of measurement items. The coordinate space of the input data is configured as an xmax-dimensional coordinate space.
The measurement items here are not limited to specific items. For example, when the graph generation device 100 generates a state change graph for predicting the state of cancer, the measurement items may be various tumor markers obtained by blood or urine sampling.

x _l (i, t _i,j ) denotes the value of the l th measurement item at time t _i,j for the i th graph generation target, where l is an integer satisfying the range 1≦l≦xmax.
j is an integer 1≦j≦tmax_i that identifies a time in the time series data. tmax_i represents the number of data items at each time included in the i-th time series data. That is, tmax_i represents the number of data measurement times in the i-th time series data. The number of data items at each time included in each time series data may differ for each time series data.

The time series data output by the feature extraction unit 210 can be expressed as a vector as shown in equation (3).

f ^* ts represents the time series data output by the feature extraction unit 210. The time series data output by the feature extraction unit 210 is also referred to as feature time series data. The feature extraction unit 210 receives input time series data fts _i and outputs feature time series data f ^* ts _i .
The feature extracted by the feature extraction unit 210 from the input data _fi (ti _,j ) at time ti _,j of the input time series data fts _i is denoted as f ^* _i (ti _,j ). f ^* _i (ti _,j ) can be expressed as a vector, for example, as shown in Equation (4).

[ _z1 (i,ti _,j ), _z2 (i,ti _,j )] is a vector representing the feature extracted by the feature extraction unit 210 from the input data [ _x1 (i,ti _,j ), _x2 (i,ti _,j ),..., _xxmax (i,ti _,j )] at time _ti, j. Equation (4) shows an example in which the feature extraction unit 210 converts xmax-dimensional input data into two-dimensional feature. _z1 (i,ti _,j ) and _z2 (i,ti _,j ) represent the coordinate value of the _z1 coordinate and the coordinate value of the _z2 coordinate, respectively.

However, as described above, the feature extraction performed by the feature extraction unit 210 is not limited to obtaining two-dimensional features.
The feature extraction unit 210 converts the input data for each time [x ₁ (i, t _{i, j} ), x ₂ (i, t i, j ), ..., x _xmax (i, t _{i, j )} ] included _in the input time series data fts i into features [z ₁ (i, t _{i, j} ), z ₂ (i, t _{i, j} )], thereby converting the input time series data fts _i into feature time series data f ^* ts _i .

The graph formation unit 220 generates a state change graph that shows the feature time series data output by the feature extraction unit 210 in a feature space. In the state change graph generated by the graph formation unit 220, each feature included in the feature time series data is plotted in the feature space, and the time series is shown by connecting temporally adjacent features in the feature time series data with directed edges.

Specifically, the graph creation unit 220 converts the data representation format of the feature time series data output by the feature extraction unit 210 so that it can search for feature time series data in which features are plotted at coordinate values in the feature space from those coordinate values.
When the graph generation device 100 generates a condition change graph to be presented to a user such as a doctor, the graph forming unit 220 may generate image data showing the condition change graph.

The learning control unit 194 controls the feature extraction unit 210 to cause the graph generation unit 192 to learn the feature extraction method described above. In this case, the learning control unit 194 may control the learning of the feature extraction unit 210 so that a specific state determined using the state change graph appears in a specific region in the feature space.

To this end, the learning control unit 194 may control the learning of the feature extraction unit 210 using an evaluation function that is highly evaluated when a feature classified into a specific class is plotted within an area associated with that class.
For example, consider a case where a feature is represented by a vector [z ₁ , z ₂ ] and the range of possible coordinate values in the feature space is 0≦z ₁ ≦1 and 0≦z ₂ ≦1. In this case, it is desired that the feature when a person is in a healthy state be plotted at or near the origin [0,0].

In this case, the learning control unit 194 may use an evaluation function that gives a higher evaluation as the distance √( _z12 + _z22 ) between the ^feature amount [ _z1 , _z2 ] and ^the origin [0, 0] when the person is healthy decreases. Alternatively, the learning control unit 194 may use an evaluation function that gives a higher evaluation when ^√ ( _z12 + _z22 ) ≤ THr holds. Here, THr is a positive real number that represents the threshold value of the distance between ^the feature amount [ _z1 , _z2 ] and the origin [0, 0].

In addition, the learning control unit 194 controls the learning of the feature extraction unit 210 so that the feature extraction unit 210 extracts features so that the length of time in the time series data is proportional to the distance in the feature space. This means that the feature space is set so that the length of time in the time series data is proportional to the distance in the feature space.

Figure 3 shows an example of a graph in which the time represented by distance in feature space varies depending on the position within the feature space. In the example of Figure 3, points P111, P112, and P113 each represent a feature. Furthermore, the time interval between the observation of data linked to point P111 (data converted into the feature represented by point P111) and the observation of data linked to point P112 is equal to the time interval between the observation of data linked to point P112 and the observation of data linked to point P113.

On the other hand, on the graph, the distance from point P112 to point P113 is longer than the distance from point P111 to point P112.
For example, consider a case where a patient's condition has changed from point P111 to point P112, and a doctor is predicting the patient's future condition by referring to the graph shown in Figure 3. In this case, if the doctor views the graph assuming that distance on the graph is proportional to time, he or she may interpret the time it takes for the patient's condition to reach point P113 as longer than it actually is. This could result in treatment for the patient being delayed.

In contrast, if a graph is generated so that distance on the graph is proportional to time, such as by plotting point P113 so that the distance between points P112 and P113 is equal to the distance between points P111 and P112 in the example of Figure 3, doctors can more accurately read the time it takes for a patient's condition to change. This is expected to enable treatment for patients to be administered at the appropriate time.

To make the distance on the graph proportional to time, the learning control unit 194 may control the learning of the feature extraction unit 210 using an evaluation index that increases the evaluation the closer the distance in feature space between two features obtained by extracting features from data at two times included in the same input time series data is to the value obtained by multiplying the time interval between those two times by a predetermined proportionality coefficient.
For example, the learning control unit 194 may control the learning of the feature extraction unit 210 using an evaluation index that increases the evaluation as the value of equation (5) decreases.

"a" is a positive real number and represents a proportionality coefficient between the distance in the feature space and the time interval in the time series data.
"||f ^* _i (t _i,j+1 )-f ^* _i (t _i,j )|| ₂ " represents the distance (L2 norm) between feature f ^* _i (t _i,j+1 ) and feature f ^* _i (t _i,j ) in the feature space, and is shown, for example, as in equation (6).

Equation (6) shows an example where the number of dimensions of the feature is two. However, as described above, the number of dimensions of the feature used in the state change graph is not limited to two. Furthermore, a distance other than the L2 norm may be used as the distance in the feature space.
The features f ^* _i (t _i,j+1 ) and f ^* _i (t _i,j ) correspond to two features obtained by extracting features _from the data f _i (t i,j+1) and f _i (t _{i ,j} ) at two times t _{i,j+1 and} t i,j, respectively, included in the same input time series data fts _i . "a(t i, _j+1 -t _i,j ) ² " corresponds to the value obtained by multiplying the time interval "t _i,j+ 1 -t i,j" between times t _i,j and t _i,j+1 in the input time series data fts _i of the two data f i (t _{i ,j+1} ) and f _i (t i _,j ) by a proportional constant "a".

Figure 4 shows a first example of the configuration of the feature extraction unit 210. In the configuration shown in Figure 4, the graph generation unit 192 includes an autoencoder (AE) 210a and a graph formation unit 220. The autoencoder 210a includes an encoder 211a, a relay layer 212a, and a decoder 213a.

The autoencoder 210a is an example of the feature extraction unit 210. Specifically, the encoder 211a receives input data for each time and outputs a feature.
The relay layer 212 a outputs the feature amount extracted by the encoder 211 a from the input data for each time to both the decoder 213 a and the graph formation unit 220 .
The decoder 213a performs data conversion on the feature quantities extracted from the input data for each time by the encoder 211a. The time series data gts _i output by the decoder 213a for the input time series data fts _i shown in equation (1) can be expressed as a vector as shown in equation (7).

g _i (t _i,j ) is data that the decoder 213a outputs upon receiving the input of the feature f _i (t _i,j ). g _i (t _i,j ) can be expressed as a vector as in equation (8).

ymax represents the number of dimensions of the output data g _i (t _i,j ) of the decoder 213a. Typically, ymax = xmax is set. In other words, the autoencoder 210a outputs data with the same number of dimensions as the number of dimensions of the input data.
In the autoencoder 210a, the data conversion of each of the encoder 211a and the decoder 213a is the subject of learning. The learning control unit 194 controls the learning of the autoencoder 210a so that the smaller the difference between the input data _fi (ti _,j ) to the encoder 211a and the output data g _i (ti _,j ) of the decoder 213a, the higher the evaluation, thereby making it possible to cause the encoder 211a and the decoder 213a to perform learning of the data conversion so that the reproducibility of the input data from the feature is improved. In this regard, the learning control unit 194 can cause the encoder 211a to perform learning so that the encoder 211a loses as little information contained in the input data as possible.
The learning control unit 194 can control the learning, particularly in the learning of the encoder 211 a, in the same manner as described above for the feature extraction unit 210.

Figure 5 is a diagram showing a second example of the configuration of the feature extraction unit 210. In the configuration shown in Figure 5, the graph generation unit 192 includes a GCM (Generative Contribution Mappings) 210b and a graph formation unit 220. The GCM 210b includes an encoder 211b, a relay layer 212b, first class decoder 213b-1 through Nth class decoder 213b-N, first multiplier 214b-1 through Nth multiplier 214b-N, first average calculation unit 215b-1 through Nth average calculation unit 215b-N, and an Argmax calculation unit 216b. Here, N is a positive integer indicating the number of classes in the classification.

The first class decoder 213b-1 through the Nth class decoder 213b-N are collectively referred to as decoders 213b. The first multiplier 214b-1 through the Nth multiplier 214b-N are collectively referred to as multipliers 214b. The first average calculation unit 215b-1 through the Nth average calculation unit 215b-N are collectively referred to as average calculation unit 215b.

The GCM 210b corresponds to an example of the feature extraction unit 210. Specifically, the encoder 211b receives input data for each time and outputs a feature.
The relay layer 212 b outputs the feature amounts extracted by the encoder 211 a from the input data for each time to each of the N decoders 213 b and the graph formation unit 220 .

Each of the N decoders 213b performs data conversion on the feature extracted from the input data for each time by the encoder 211b to convert it into data with the same number of dimensions as the number of dimensions of the input data.
The multiplier 214b-i (i = 1, 2, ..., N) multiplies the data calculated by the i-th class decoder 213b-i by the input data. The data output by the multiplier 214b-i can be used as information indicating how much each part of the input data contributes to the determination of the i-th class.

The average calculation unit 215b-i calculates the average of all element values of the data calculated by the multiplier 214b-i. That is, the average calculation unit 215b-i calculates the average of the values for each dimension of the xmax-dimensional data calculated by the multiplier 214b-i. The average values calculated by each average calculation unit 215b are used as evaluation values (class scores) in class classification.
The Argmax calculation unit 216b compares the class scores calculated for each class by the average calculation unit 215b and determines the class with the largest class score. In this way, the Argmax calculation unit 216b classifies the input data into classes.

In GCM 210b, encoder 211b and N decoders 213b are the targets of learning. The learning control unit 194 can control learning, particularly in the learning of encoder 211b, in the same manner as described above for feature extraction unit 210.

The method by which the feature extraction unit 210 extracts features is not limited to the method using the autoencoder 210a or the method using the GCM 210b. The feature extraction unit 210 can use various methods that can convert input data into data of the desired dimension as a method for extracting features.

The state information addition unit 193 adds information to a portion of the feature space (the coordinate space in which the state change graph is displayed) indicating the state of the prediction target at the time in the input time series data of the data converted into coordinate values included in that portion. For example, the state information addition unit 193 may add the name of the state indicated by the feature to an area including a point at which the feature is plotted.

Furthermore, for example, as described above, when the learning control unit 194 controls the learning of the feature extraction unit 210 so that a specific state appears in a specific region, the state information addition unit 193 may add the name of that state to that region. In the above example, the state information addition unit 193 may add the state name "healthy" to the region within a radius THr from the origin [0,0].

Figure 6 is a diagram showing an example of a graph output by the graph generation device 100. Figure 6 shows an example of changes in the condition of a hospitalized patient due to a certain illness. Each node in the graph in Figure 6 represents the patient's condition. Specifically, the coordinate values of the node represent features extracted from the measurement data about the patient. Point P211 represents the state when the patient's condition has improved and led to discharge from the hospital. Point P212 represents the state when the patient's condition has worsened and led to death.

An edge represents the temporal relationship between two adjacent features in the feature time-series data. In Figure 6, the edge direction is indicated by an arrow only for line L11 from point P221 to P212, but the edge direction for other edges may also be as shown in the figure.

In addition, in Figure 6, the length of each edge is proportional to the time interval in the feature time-series data between the two feature amounts represented by the nodes at both ends of the edge. The closer the distance on the graph (i.e., the distance when tracing the edge) to point P211, the better the patient's condition can be determined to be. On the other hand, the closer the distance on the graph to point P212, the worse the patient's condition can be determined to be.

For example, a user such as a doctor who wants to predict a patient's condition uses the graph generation device 100 to extract features from data indicating the current condition of the patient whose condition is to be predicted, and plots the obtained features on a graph. For example, by tracing the edges from the point closest to the plotted point, the user can predict how the patient's condition will change and over what period of time.

Figure 7 shows a first example of states indicated by a state change graph. In a state change graph, one node plotted as a point in coordinate space represents one state. In the example of Figure 7, points P311, P312, P321, and P322 each represent one state. In the following description, there may be cases where no distinction is made between a node and the point at which that node is plotted in state space.

Furthermore, nodes that indicate similar states are placed so that the distance in coordinate space is small (i.e., close by). Nodes that indicate dissimilar states are placed so that the distance in coordinate space is large (i.e., farther away). In the example of Figure 7, the distance d11 between points P311 and P312 is relatively small, and the distance d12 between points P321 and P322 is relatively large. From this, it can be determined that the state represented by point P311 and the state represented by point P312 are relatively similar. It can be determined that the state represented by point P321 and the state represented by point P322 are relatively different.

Figure 8 is a diagram showing a second example of states indicated by a state change graph. The edges of a state change graph can be understood as vectors that represent changes in a state over time. In the example of Figure 8, for example, the edge from point P411 to point P412 can be understood as a vector that indicates the change from the state represented by point P411 to the state represented by point P412. Furthermore, the state change graph can be read as the length of this edge being proportional to the time it takes to change from the state represented by point P411 to the state represented by point P412.

Also, a certain node may serve as the starting point in a graph. In the example of Figure 8, point P411 is the starting node in the graph of Figure 8. The starting node is linked to the first data (data at the oldest time) in the input time series data, and it can be considered that there is no measurement data at a time earlier than that.

In addition, a certain node may be the end point in a graph. In the example of Figure 8, point P414 is the end point node in the graph of Figure 8. The start point node is linked to the last data (data at the most recent time) in the input time series data, and it can be considered that there is no measurement data at a time more recent than that.

In addition, a node may become an intermediate node in the graph. An intermediate node represents a state that has changed from another state, and can be understood as a state that has changed to yet another state. In the example of Figure 8, the state represented by point P412 is a state that has changed from the state represented by point P411. Furthermore, the state represented by point P412 has changed to the state represented by point P413.

FIG. 9 is a diagram showing a third example of a state indicated by a state change graph.
When multiple feature quantities are plotted at the same position in the feature quantity space and represented by a single node, multiple edges may emerge from a single node. In the example of Figure 9, point P511 is considered to represent the most recent feature quantity of the feature quantity indicated by point P512 on the feature quantity time-series data, and the most recent feature quantity of the feature quantity indicated by point P514 on the feature quantity time-series data. For this reason, an edge from point P511 to point P512 and an edge from point P511 to point P514 are established.

In addition, when multiple feature quantities are plotted at the same position in feature space and represented by a single node, multiple edges may reach a single node. In the example in Figure 9, point P514 is thought to represent the near-future feature quantity of the feature quantity indicated by point P513 on the feature time-series data, and the near-future feature quantity of the feature quantity indicated by point P511 on the feature time-series data. For this reason, an edge is established from point P513 to point P514, and an edge is established from point P511 to point P514.

Also, there may be multiple paths from a node to another node. In such cases, the state change graph can be interpreted by assuming that the sum of the lengths of the edges included in each path is proportional to the time required for a state change when following that path. In the example of Figure 9, the sum of the lengths of the edges on the path from point P511 to point P514 via points P12 and P513 is 3. On the other hand, the path from point P511 directly to point P514 has one edge, and its length is 2. From this, we can interpret the time required for a state change from the state indicated by point P511 to the state represented by point P514 via the states indicated by points P512 and P513 as 1.5 times the time required for a state change from the state indicated by point P511 to the state represented by point P514 directly.

FIG. 10 is a diagram showing a fourth example of a state indicated by a state change graph.
In the example of Figure 10, there is an edge from point P611 to point P612 and an edge to point P613. The edge from point P611 to point P612 is shown with a probability of passing through this edge of 70%. The edge from point P611 to point P613 is shown with a probability of passing through this edge of 30%. In this way, when multiple edges extend from one node, the probability of a state change occurring through that edge may be shown for each edge.

In this case, the probability may indicate the percentage of times that, in all feature time series data, the feature associated with the node that is the starting point of the edge is the most recent feature associated with each node that is the destination of the edge.
In the example of Figure 10, when the feature values in the immediate future of the feature values linked to point P611 are counted for all feature time-series data, the feature value linked to point P612 may appear seven times, and the feature value linked to point P613 may appear three times.

FIG. 11 is a diagram showing a fifth example of a state indicated by a state change graph.
11, there are two paths from point P711 to point P714, and the path from point P711 to point P714 via point P712 has a probability of 70%, while the path from point P711 directly to point P714 has a probability of 30%.
Furthermore, there is only one path from point P713 to P714, which is a path that goes directly from point P713 to P714, and this path is shown to have a probability of 100%.

In this way, for each path from one node to another, the probability of passing through that path may be indicated.
In this case, the probability may indicate the probability of the number of times a state change occurs from the state represented by the node at the start of the path to the state represented by the node at the end of the path, and the number of times each path is taken.
In the example of Figure 11, the number of times that the state indicated by point P711 has progressed to the state indicated by point P714 via the state indicated by point P712 is seven, and the number of times that the state indicated by point P711 has progressed directly to the state indicated by point P714 is three.

FIG. 12 is a diagram showing an example of a state change graph in which the display area of an edge is determined according to the direction of the edge.
12 shows the change in the condition of a patient with a certain disease. The feature space of this condition change graph is formed as a two-dimensional space (i.e., a plane) with _z1 and _z2 axes.

Furthermore, the lower left region of the feature space, which is within a predetermined distance from the origin, is set as a region indicating a state in which the patient's condition has improved and led to discharge from the hospital. The upper right region of the feature space, which is within a predetermined distance from the maximum value of _z1 and _z2, is set as a region indicating a state in which the patient's condition has worsened and led to death. In this condition change graph, the edge from the lower left (toward the origin) to the upper right (toward the maximum values of _z1 and _z2 ) of the graph is considered to indicate a worsening of the condition. The edge from the upper right (toward the maximum values of _z1 and _z2 ) to the lower left (toward the origin) of the graph is considered to indicate an improvement of the condition.

In this case, the graph generation unit 192 may generate a state change graph in which edges indicating a worsening state and edges indicating an improvement in state are shown in separate regions. In the example of Fig. 12, edges indicating a worsening state are shown on the upper left side of the figure, and edges indicating an improvement in state are shown on the lower right side, with line L21 being the diagonal line connecting the origin and the maximum points of _z1 and _z2 as the boundary.

To this end, the learning control unit 194 may control the learning of the graph generation unit 192 using an evaluation function that gives a higher evaluation the smaller the value of equation (9).

eval is a function that outputs a value of 1 if the change from feature f(t _i,j+1 ) to f(t _i,j ) indicates a worsening of the condition, and outputs a value of −1 if the change indicates an improvement of the condition.
For example, with the positive direction of the _z1 axis as the reference (0 degrees), the value of eval may be set to 1 when the angle θ of the vector f(t _i,j+1 )-f(t _i,j ) is 0≦θ<135° and 315°≦θ<360°, and the value of eval may be set to −1 when 135°≦θ<315°.
Step represents a step function, which is expressed as in equation (10).

The value of "step(z ₁ (i, t _{i, j} ) - z ₂ (i, t _{i, j} ))" is 1 if z ₁ (i, t _{i, j} ) ≥ z ₂ (i, t _{i, j} )), and is 0 if z ₁ (i, t _{i, j} ) < z ₂ (i, t _{i, j} )). Therefore, the value of "step(z ₁ (i, t _{i, j} ) - z ₂ (i, t _{i, j} ))" is 1 if the feature quantity _fi (t _{i, j} ) = [z ₁ (i, t _{i, j} ), z ₂ (i, t _{i, j} )] is located to the lower right of line L21, and is 0 if it is located to the upper left of line L21.

The value of "step(z ₁ (i, t _{i, j+1} ) - z ₂ (i, t _{i, j+1} ))" is 1 when z ₁ (i, t _{i, j+1} ) ≥ z ₂ (i, t _{i, j+1} )), and is 0 when z ₁ (i, t _{i, j+1} ) < z ₂ (i, t _{i, j+1} )). Therefore, the value of "step(z ₁ (i, t _{i, j+1} ) - z ₂ (i, t _{i, j+1} ))" is 1 when the feature f _i (t _{i, j+1} ) = [z ₁ (i, t _{i, j+1} ), z ₂ (i, t _{i, j+1} )] is located to the lower right of line L21, and is 0 when it is located to the upper left of line L21.

Therefore, the value of "step(z ₁ (i,t _i,j+1 ) - z ₂ (i,t _i,j+1 )) + step(z ₁ (i,t _i,j+1 ) - z ₂ (i,t _i,j+1 )) - 1" is 1 when both features f _i (t _i,j ) and f _i (t _i,j+1 ) are located to the lower right of line L21, 0 when only one of them is located to the lower right of line L21, and -1 when both are located to the upper left of line L21.

If the change from feature f(t _i,j+1 ) to f(t _i,j ) indicates a worsening of the condition, the value of eval is 1, and if both feature f _i (t _i,j ) and f _i (t _i,j+1 ) are located above and to the left of line L21, the value of equation (9) becomes the minimum value 0.
If the change from feature f(t _i,j+1 ) to f(t _i,j ) indicates an improvement in condition, the value of eval is −1, and if feature f _i (t _i,j ) and f _i (t _i,j+1 ) are both located to the lower right of line L21, the value of equation (9) becomes the minimum value 0.

As a result, it is expected that an edge indicating a worsening condition will be shown on the upper left side of the feature space (area _z1 < _z2 ), and an edge indicating an improving condition will be shown on the lower right side (area _z1 ≥ _z2 ).

FIG. 13 is a diagram showing an example of a state change graph displayed in a three-dimensional feature space. In the example of FIG. 13, the feature space is configured with three coordinate axes, i.e., the _z1 axis, the _z2 axis, and the _z3 axis, and the state change graph is displayed three-dimensionally in the feature space. By displaying the state change graph three-dimensionally in this way, it is expected that differences in state will be displayed more clearly than when the state change graph is displayed two-dimensionally. For example, features that are displayed close to each other in the two-dimensional feature space may be displayed relatively far apart in the three-dimensional feature space.

FIG. 14 is a flowchart showing an example of a procedure for the graph generation device 100 to perform learning of the graph generation unit 192.
14 , the data acquiring unit 191 acquires learning data (step S111). The data acquiring unit 191 acquires, as the learning data, one or more pieces of input time-series data and information indicating the state of the graph generation target at that time, which is linked to one or more pieces of input data for each time of the input time-series data. The information indicating the state of the graph generation target is used as a correct answer (teaching data) for classifying the state of the graph generation target.

Next, the graph generation unit 192 calculates feature amounts and output data (step S112). Specifically, the graph generation unit 192 extracts feature amounts for each piece of input data for each time in the input time-series data. The feature amounts indicate coordinate values in a feature space and can be treated as nodes in a state change graph. The graph generation unit 192 then draws edges between the feature amounts as nodes in accordance with the order in which the input data appear in the input time-series data.
The graph generating unit 192 performs the process of step S112 on all input time-series data provided as learning data or on input time-series data selected by the learning control unit 194, and generates a state change graph.

Next, the learning control unit 194 determines whether a learning termination condition is met (step S113). The termination condition here is not limited to a specific one. For example, the termination condition here may be that the evaluation value of the state change graph or a part of it generated in step S112 is equal to or greater than a predetermined threshold. Alternatively, the termination condition here may be that the loop of steps S112, S113, and S114 has been repeated a predetermined number of times or more.

If the learning control unit 194 determines that the termination condition is not satisfied (step S113: NO), the graph generation unit 192 updates the learning model parameter values under the control of the learning control unit 194 (step S114). A known method may be used to update the learning model parameter values. For example, if the feature extraction unit 210 is configured using a neural network, the feature extraction unit 210 may update the parameter values of the neural network based on backpropagation under the control of the learning control unit 194.
After step S114, the process returns to step S112.
On the other hand, if the learning control unit 194 determines in step S113 that the termination condition is met (step S113: YES), the graph generation device 100 terminates the processing in FIG.

The state information adding unit 193 may add information indicating the state of the graph generation target to the state change graph finally obtained during learning, to create a state change graph to be used for state prediction.
Alternatively, a state change graph may be generated using the graph generation device 100 after learning is completed. In this case, the graph generation unit 192 may generate a state change graph by performing the same process as in step S112, and the state information addition unit 193 may add information indicating the state of the graph generation target to the obtained state change graph.

As described above, the graph generation unit 192 converts each piece of data (input data for each time) included in the time series (input time series data) of data related to the prediction target (graph generation target) into coordinate values in a coordinate space set so that the length of time in the time series is proportional to the distance in the coordinate space, and generates a graph in the coordinate space (state change graph). The state information addition unit 193 adds, to a portion of the coordinate space, information indicating the state of the prediction target at the time in the time series (input time series data) of the data (input data for each time) converted into coordinate values included in that portion.

The graph generating device 100 can visualize a predicted state change and show the relative length of time it takes for the state change to occur.
Specifically, the graph generation device 100 generates a state change graph that shows how the state of the graph generation target changes and includes information indicating the state, such as the name of the state. A person or a state prediction device can plot the current state of the prediction target on the state change graph to predict the future state.
In this case, since the state change graph is generated so that the distance in the coordinate space of the state change graph is proportional to the time interval in the input time series data, it is expected that a person or a state prediction device will be able to predict the time required for a state change with relatively high accuracy.

The data acquisition unit 191 acquires input time-series data. The coordinate space of the state change graph can be regarded as a feature space. The graph generation unit 192 converts each piece of data included in the time series into a feature using a feature extraction method obtained by learning using an evaluation index in which the closer the distance in the feature space between two feature amounts obtained by extracting feature amounts from each of two pieces of data (input data for each time point) included in the same time series (input time-series data) is to a value obtained by multiplying the time interval between the two pieces of data in the time series by a predetermined proportional coefficient, the higher the evaluation.
In the graph generation device 100, the process of calculating the coordinate values of nodes in a state change graph can be performed as a feature extraction process. In this respect, the graph generation device 100 can use a feature extraction method to calculate the coordinate values of nodes in a state change graph. A known feature extraction method may be used to calculate the coordinate values of nodes in a state change graph.

Second Embodiment
When generating a state change graph after learning is completed (during operation), the configuration for learning may be removed from the configuration of the graph generation device 100. This point will be described in the second embodiment.
Fig. 15 is a diagram illustrating an example of the configuration of a graph generation device according to the second embodiment, when generating a state change graph after learning is completed.
15 , the graph generation device 100b includes a communication unit 110, a display unit 120, an operation input unit 130, a storage unit 180, and a control unit 190. The control unit 190b includes a data acquisition unit 191, a graph generation unit 192, and a state information addition unit 193.

15, parts having the same functions as those in FIG. 1 are denoted by the same reference numerals (110, 120, 130, 180, 191, 193), and detailed description thereof will be omitted here. The graph generation device 100b differs from the control unit 190 of the graph generation device 100 in that the control unit 190b does not include a learning control unit 194. In other respects, the graph generation device 100b is similar to the graph generation device 100.
When generating a state change graph, the graph generation device 100 may be used in the same way as when learning. Alternatively, a graph generation device 100b may be prepared separately from the graph generation device 100, and the parameter values of the learning model of the graph generation unit 192 obtained by the graph generation device 100 may be set in the graph generation unit 192 of the graph generation device 100b.

Fig. 16 is a diagram showing a third example of the configuration of the feature extraction unit 210. Fig. 16 shows an example of the configuration of the feature extraction unit 210 when generating a state change graph in a case where the feature extraction unit 210 is configured using an autoencoder during learning.
16 , the graph generation unit 192 includes a feature extraction unit 210 and a graph formation unit 220. The feature extraction unit 210 includes an encoder 211 and a relay layer 212.

Of the components in Figure 16, those that have similar functions to those in Figure 4 are given the same reference numerals (211a, 212a, 220) and will not be described in detail here. The configuration of the feature extraction unit 210 in Figure 15 is the same as the configuration of the autoencoder 210a in Figure 4, except that the decoder 213a is removed. In this way, if a configuration for learning is added to the feature extraction unit 210 during learning, that configuration is not necessary when generating a state change graph.

When using the graph generation device 100 to generate a state change graph, the feature extraction unit 210 may also use the autoencoder 210a in Figure 4, as in the case of learning. Alternatively, when the graph generation device 100b is prepared separately from the graph generation device 100, the feature extraction unit 210 may have the configuration shown in Figure 16. In this case, the parameter values of the learning model of the encoder 211a obtained during learning may be set in the encoder 211a used when generating a state change graph.

Third Embodiment
In the third embodiment, a state prediction using a state change graph will be described.
FIG. 17 is a diagram illustrating an example of the configuration of a state prediction system according to the third embodiment.
In the configuration shown in FIG. 17, the state prediction system 1 includes a graph generation device 100 and a state prediction device 300.

The graph generation device 100 in the first embodiment can be used as the graph generation device 100. Alternatively, instead of the graph generation device 100, the graph generation device 100b in the second embodiment or the graph generation device 400 in the fourth embodiment described below may be used.
The state prediction device 300 predicts the state of the prediction target using the state change graph generated by the graph generation device 100. The state prediction device 300 may be configured using a computer such as a workstation or a personal computer.

FIG. 18 is a diagram illustrating an example of the configuration of a state prediction device 300.
18 , the state prediction device 300 includes a communication unit 310, a display unit 320, an operation input unit 330, a storage unit 380, and a control unit 390. The control unit 390 includes a data acquisition unit 391 and a prediction unit 392.

The communication unit 310 communicates with other devices. For example, the communication unit 310 may communicate with the graph generation device 100 to receive a state change graph. The communication unit 310 may also communicate with the graph generation device 100 to obtain information indicating learning model parameter values obtained by learning by the graph generation device 100. The communication unit 310 may also communicate with a measurement device that measures data related to the prediction target to obtain the latest measurement data related to the prediction target.

The display unit 320 has a display screen such as a liquid crystal panel or an LED panel, and displays various images. For example, the display unit 320 may display a state change graph and a state prediction result of a prediction target using the state change graph.
The operation input unit 330 includes input devices such as a keyboard and a mouse, and receives user operations. For example, the operation input unit 330 may receive a user operation to instruct the start of state prediction and a user operation to instruct a source for obtaining data related to the prediction target.

The storage unit 380 stores various data. For example, the storage unit 380 stores the state change graph obtained from the graph generation device 100. The storage unit 380 is configured using a storage device included in the state prediction device 300.
The control unit 390 performs various processes by controlling each unit of the state prediction device 300. The functions of the control unit 390 are performed, for example, by a CPU included in the state prediction device 300 reading and executing a program from the storage unit 380.

The data acquisition unit 391 acquires various data for the state prediction device 300 to predict the state of the prediction target. In particular, the data acquisition unit 391 acquires a state change graph and data related to the state of the prediction target.
When the communication unit 310 receives a state change graph from the graph generation device 100, the data acquisition unit 391 may read the state change graph from the data received by the communication unit 310. Alternatively, when the storage unit 380 stores a state change graph in advance, the data acquisition unit 391 may read the state change graph from the storage unit 380.
Furthermore, when the communication unit 310 receives data related to the state of the prediction target from another device, the data acquisition unit 391 may read the data related to the state of the prediction target from the data received by the communication unit 310. It is preferable that the data acquisition unit 391 acquires data of items similar to the items of input data for each time in the input time-series data when the graph generation device 100 creates a state change graph.

The prediction unit 392 predicts the state of the prediction target using the state change graph.
Specifically, feature quantities are extracted from data related to the state of the prediction target. In particular, the prediction unit 392 extracts feature quantities using the same feature quantity extraction method as that used by the feature quantity extraction unit 210 of the graph generation device 100. As a result, the feature quantities extracted by the prediction unit 392 represent coordinate values in a feature quantity space, which is the coordinate space of the state change graph. Therefore, the feature quantities extracted by the prediction unit 392 can be treated as plotting the feature quantities on the state change graph.
Then, the prediction unit 392 determines which node of the state change graph to refer to based on the extracted feature amount, and predicts the state of the prediction target based on the edge going out from the determined node.
Hereinafter, the feature extracted by the prediction unit 392 will be referred to as a feature to be predicted.

Fig. 19 is a diagram showing an example of nodes in the state change graph that are referenced by the state prediction device 300. In the example of Fig. 19, points P811 and P821 each indicate an example of displaying a prediction target feature in the state change graph.
Point P811 shows an example in which the prediction unit 392 selects one node from the nodes in the state change graph. When there is one or more nodes that are within a predetermined first distance threshold from the feature to be predicted, the prediction unit 392 may select one of these nodes that is closest to the feature to be predicted. Area A11 corresponds to an example of an area within the first distance threshold from the feature to be predicted. Furthermore, point P812 corresponds to an example of a node selected by the prediction unit 392.

In this case, since the position of the feature to be predicted is close to the position of the node selected by the prediction unit 392, there is a relatively high possibility that the state of the target to be predicted will change in a manner similar to the state change indicated by the edge from the node selected by the prediction unit 392.
Therefore, the prediction unit 392 may trace edges going out from the selected node, and if information indicating a state is attached to the reached node, predict that state as the future state of the prediction target. Alternatively, the prediction unit 392 may multiply the length of the traced edge by a predetermined proportionality coefficient to calculate a predicted value of the time required for a state change. In this case, the proportionality coefficient may be the reciprocal (1/a) of the proportionality coefficient "a" in equation (5).
The prediction unit 392 may repeat tracing edges from the destination node to predict a further future state of the prediction target.

On the other hand, point P821 shows an example in which the prediction unit 392 selects one or more nodes from the nodes of the state change graph. If there are no nodes within a predetermined first distance threshold from the feature to be predicted, the prediction unit 392 may select all nodes within a second distance threshold that is longer than the first distance threshold. Area A21 is an example of an area within the second distance threshold from the feature to be predicted. Furthermore, points P822, P823, and P824 are examples of nodes selected by the prediction unit 392.

In this case, the prediction unit 392 may calculate an intermediate edge among the edges going out from the selected node and treat it as a prediction of the state change of the prediction target.
For example, the prediction unit 392 may treat edges as vectors in feature space and calculate a weighted average of the vectors according to the distance from the feature to be predicted to the node (the starting point of the vector).The prediction unit 392 may then treat the position advanced by the vector calculated from the feature to be predicted as a prediction of the future state of the prediction target.In this case, the prediction unit 392 may also multiply the length of the calculated vector by a predetermined proportional coefficient to calculate a prediction value of the time required for the state change.
The prediction unit 392 may also predict a further future state of the prediction target.

FIG. 20 is a flowchart illustrating an example of processing by the state prediction device 300 to predict the state of a prediction target.
20 , the data acquisition unit 391 acquires data for predicting the state of the prediction target (step S201). For example, the data acquisition unit 391 acquires a state change graph and data related to the state of the prediction target. The data acquisition unit 391 acquires data of items similar to the items of input data for each time in the input time-series data when the graph generation device 100 creates a state change graph.

Next, the prediction unit 392 extracts features of data related to the state of the prediction target (step S202). The prediction unit 392 extracts features using the same feature extraction method as in the case of the feature extraction unit 210 of the graph generation device 100.
Next, the prediction unit 392 selects a node in the state change graph (step S203). As described with reference to Fig. 19, the prediction unit 392 may select one node that is closest to the feature to be predicted, or may select one or more nodes within a predetermined distance from the feature to be predicted.

Next, the prediction unit 392 predicts the state of the prediction target based on the edges outgoing from the selected node (step S204). As described with reference to Fig. 19, the prediction unit 392 may treat the node reached by tracing an edge outgoing from one node selected in step S203 as a prediction of the future state of the prediction target. Alternatively, the prediction unit 392 may treat the edges outgoing from one or more nodes selected in step S203 as a vector, perform a weighted average, and treat the position advanced from the feature quantity of the prediction target by the amount of the obtained vector as a prediction of the future state of the prediction target.
The prediction unit 392 may repeat the processes of steps S203 and S204 to predict a further future state of the prediction target.
After step S204, the state prediction device 300 ends the process of FIG.

As described above, the prediction unit 392 predicts the state of the object to be predicted using a graph in which the time series of data related to the object to be predicted is shown in a coordinate space set so that the length of time in the time series is proportional to the distance in the coordinate space.
According to the state prediction device 300, it is possible to calculate a predicted value of the time required for a state change by a relatively simple calculation, for example, by multiplying the length of an edge in the state change graph by a predetermined proportionality coefficient.

In addition, the prediction unit 392 detects data (nodes of the state change graph) at which the state of the prediction target before the state change, indicated as the state prediction target, is plotted in the coordinate space of the state change graph and is the shortest distance in that coordinate space, and predicts the state of the prediction target after the state change based on the edges of the state change graph from the detected data to other data.
In this case, the prediction unit 392 can select only one node in the state change graph. In this respect, the prediction unit 392 can perform state prediction with relatively simple processing, such as treating a node reached by tracing an edge from the selected node as a future state prediction for the prediction target.

In addition, the prediction unit 392 detects data (nodes of the state change graph) located within a predetermined distance in the coordinate space from the point at which the state of the prediction target before the state change, which is indicated as the state prediction target, is plotted in that coordinate space, and predicts the state of the prediction target after the state change based on the edges of the state change graph from the detected data to other data.
This allows the prediction unit 392 to predict the state of the prediction target. In particular, by selecting multiple nodes, the prediction unit 392 can perform state prediction with relatively high accuracy, for example, by calculating a weighted average of the edges coming out from the multiple nodes.

<Fourth embodiment>
In the fourth embodiment, a case will be described in which a graph generation device generates a state change graph by converting an already generated graph.
FIG. 21 is a diagram illustrating an example of the configuration of a graph generation device according to the fourth embodiment.
21 , the graph generation device 400 includes a communication unit 410, a display unit 420, an operation input unit 430, a storage unit 480, and a control unit 490. The control unit 490 includes a data acquisition unit 491, a graph generation unit 492, and a state information addition unit 493.

The graph generation device 400 converts the graph. In particular, the graph generation device 400 converts a graph in which the ratio between the length of the graph edges and the time in the time-series data varies depending on the position in the coordinate space, so that the ratio between the length of the graph edges and the time in the time-series data becomes constant. Through this conversion, the graph generation device 400 generates a state change graph.
The graph generating device 400 may be configured using a computer such as a workstation or a personal computer.

The communication unit 410 communicates with other devices. For example, the communication unit 410 may communicate with a device that has generated a graph and receive the graph.
The display unit 420 has a display screen such as a liquid crystal panel or an LED panel, and displays various images. For example, the display unit 420 may display a graph before conversion and a graph after conversion.
The operation input unit 430 includes input devices such as a keyboard and a mouse, and receives user operations. For example, the operation input unit may receive a user operation to instruct graph conversion and a user operation to instruct a source from which to obtain a graph.

The storage unit 480 stores various data. For example, the storage unit 480 may store a graph before conversion and a graph after conversion.
The control unit 490 performs various processes by controlling each unit of the graph generation device 400. The functions of the control unit 490 are performed, for example, by the CPU included in the graph generation device 400 reading and executing a program from the storage unit 480.

The data acquisition unit 491 acquires, as the graph before conversion, a graph in which the ratio of the length of the edge of the graph to the time in the time-series data varies depending on the position in the coordinate space.
For example, when the communication unit 410 receives a graph from another device, the data acquisition unit 491 may read the graph from the signal received by the communication unit 410 .

The graph generation unit 492 performs graph transformation. As described above for the graph generation device 400, the graph generation unit 492 performs transformation on a graph in which the ratio between the length of the graph edges and the time in the time series data varies depending on the position in coordinate space, so that the ratio between the length of the graph edges and the time in the time series data becomes constant.

The state information addition unit 493 adds information indicating the state of the prediction target at the time in the input time series data of the data converted into coordinate values included in a portion of the coordinate space of the graph generated by the graph generation unit 492. The state information addition unit 493 may detect information added to the graph acquired by the data acquisition unit 491, and add the detected information to an area of the converted graph corresponding to the area to which the detected information has been added.

Fig. 22 is a diagram showing an example of a graph before conversion. In the example of Fig. 22, each node of the graph represents a feature amount of data obtained by repeatedly measuring the graph generation target at a fixed time interval. Meanwhile, in the graph shown in Fig. 22, the length of the edges varies depending on the position in the coordinate space. Therefore, in the graph shown in Fig. 22, the ratio of the length of the graph edges to the time in the time-series data varies depending on the position in the coordinate space.
The time series data f ^m ts _i represented by the graph before conversion can be expressed as in equation (11).

As in the case of equation (3), in equation (11), tmax_i represents the number of data for each time included in the i-th time series data.
In the example of FIG. 22, the number of time series data is k (k is a positive integer), and is expressed as f ^m ts ₁ , f ^m ts ₂ . . . , f ^m ts _k .
Data f ^m (t _i,j ) for each time in the time series data f ^m ts _i is expressed as in equation (12).

[ ^zm1 (t _i,j ), ^zm2 (t _i,j )] indicates the coordinate values of _a node of the graph in the two-dimensional coordinate space formed by _{the zm1} ^coordinate axis and _the ^zm2 coordinate axis. Hereinafter, _the node represented by data ^fm _i (t _i,j ) will also be referred to as node ^fm _i (t _i,j ).
Fig. 23 is a diagram showing an example of a graph after conversion: Fig. 23 shows an example of a graph obtained by converting the graph of Fig. 22.
In the graph of FIG. 22, the lengths of the edges of the graph vary depending on the position in the coordinate space, whereas in the graph of FIG. 23, the lengths of the edges are approximately constant.

When the graph generation unit 492 converts a graph, if an area associated with a specific status category, such as "health," is set in the original graph, it is preferable to set an area equivalent to that area in the converted graph as well. To achieve this, it is possible to reflect the distance between nodes in the original graph in addition to the time interval in the time series data in the distance between nodes in the converted graph, so that two nodes that are close to each other in the original graph will also be close to each other in the converted graph.

For example, the graph generating unit 492 may set the distance D between nodes in the converted graph as follows, and repeat the conversion of the graph.
For two nodes connected by an edge, the graph generation unit 492 calculates and sets the distance between the nodes based on equation (13).

As shown in the examples of nodes f ^m ₂ (t _2,2 ) and f ^m ₂ (t _2,3 ) in FIG. 22, the data of nodes f ^m _i (t _i,j ) and f ^m _i (t _i,j+1 ) are adjacent in time on the time series, and these two nodes are connected by an edge.
The values before the current conversion are used as the values of ^fm ₍ tj _,j ) and ^fm (tj _{,j+1 )} , which represent the coordinate values of the nodes. For example, in the first conversion, the values in the graph acquired by the data acquisition unit 491 are used as the values of ^fm ₍ tj _,j ) and ^fm _{( tj,j+1} ). In the second and subsequent conversions, the values in the graph obtained in the conversion immediately before that are used.
"||f ^m _i (t _i,j+1 ) - ^{f m} _i (t _i,j )|| ₂ " represents the distance (L2 norm) in the coordinate space before the current transformation between two nodes f ^m _i (t _i,j ) and f ^m _i (t _i,j+1 ) connected by an edge.

"t _i,j+1 -t _i,j " represents the time interval between nodes f _i ^m (t _i,j ) and f _i ^m (t _i,j+1 ) in the time series data.
Each of α and β is a positive real constant. Alternatively, the value of α and/or β may be changed, such as by halving the value of α after a predetermined number of iterations of the transformation.

In this way, by reflecting the distance between nodes based on the distance before the current change and the time interval in the time series data, the distance between nodes in the graph acquired by the data acquisition unit 491 can be reflected, and as the conversion is repeated, the distance between nodes can gradually approach a distance proportional to the time interval in the time series data.

For two nodes that are not connected by an edge, the graph generation unit 492 calculates and sets the distance D between the nodes based on equation (14).

As in the example of nodes f ^m ₂ (t _2,2 ) and f ^m _k (t _k,1 ) in FIG. 22, it is assumed that node f ^m _i (t _i,j ) and node f ^m _i′ (t _i′,j′ ) are not connected by an edge.
γ and K are real constants, or the values of γ and/or K may be varied.
In equation (14), the two nodes are not connected by an edge, and the concept of time for a state change does not apply between these two nodes, so a time interval term such as "β×(t _i,j+1 −t _i,j )" in equation (13) is not provided. On the other hand, equation (14) provides a term for the distance in the graph before conversion, "γ×∥f ^m _i (t _i,j ) − ^{f m} _i' (t _i',j' )∥ ₂ ."

The method by which the graph generation unit 492 generates a converted graph so that the distance setting is reflected is not limited to a specific method. For example, the graph generation unit 492 may set any one node as a reference node and determine the position of each node in order from the reference node, reflecting the set distance. If graph generation fails, the graph generation unit 492 may reset the reference node to another node and redo the process of determining the node positions in order.
When the graph generation unit 492 converts a graph, a constraint or evaluation function may be set to minimize the distance that the nodes move before and after the conversion, so as to maintain the positional relationship of the nodes on the graph acquired by the data acquisition unit 491 as much as possible.

Alternatively, the set distance does not necessarily have to be reflected accurately for all pairs of two nodes for which a distance is set. For example, a constraint may be set such that the closer the set distance (the smaller the distance value), the closer the distance in the transformed graph will be, and an error between the set distance under the set constraint and the distance in the transformed graph may be tolerated.

The graph acquired by the data acquisition unit 491 corresponds to an example of a first graph in which a time series of data related to the prediction target is shown in a first coordinate space. The time series data f ^m ts _i corresponds to an example of a time series of data related to the prediction target.
The distance D calculated by equation (13) is an example of a distance index that is a weighted sum of the distance in the first coordinate space between two pieces of data included in the time series shown in the first coordinate space and the time interval in the time series of these two pieces of data.

The conversion performed by the graph generation unit 492 is an example of a coordinate conversion from first coordinates to second coordinates such that the smaller the value of the distance index, the closer the distance between the two data in the second coordinate space.
The graph generated by the graph generating unit 492 through the conversion corresponds to an example of a second graph in which the time series of data is shown in a second coordinate space.

The data acquisition unit 491 may acquire the graph to be converted by the graph generation unit 492 (the first graph described above) based on tidal flow diagram data, which is data indicating state change rate vectors at each position in coordinate space.

FIG. 24 is a diagram showing a first example in which the graph generating device 400 generates a graph from a power flow diagram of a state change speed vector.
24, the data acquisition unit 491 sets sampling points at regular intervals within a coordinate space in which the state change speed vector is indicated. The data acquisition unit 491 sets one of the set sampling points as a starting point, sets a sampling point at a position a predetermined time ahead of the starting point according to the state change speed vector, and sets a directed edge from the starting point to the newly set sampling point.

Furthermore, the data acquisition unit 491 sets the newly set sampling point as the starting point, sets a sampling point at a position a predetermined time ahead from the starting point according to the state change rate vector, and sets a directed edge from the starting point to the newly set sampling point.
The data acquisition unit 491 repeats the setting of new sampling points and directed edges by advancing in accordance with the state change speed vector, for example, a predetermined number of times, or until the advancing destination goes outside the domain of the coordinate space.

The data acquisition unit 491 repeatedly sets new sampling points and directed edges by moving forward from each of the sampling points set at regular intervals in accordance with the state change rate vector. In this way, the data acquisition unit 491 acquires a graph.

FIG. 25 is a diagram showing a second example in which the graph generating device 400 generates a graph from a power flow diagram of a state change speed vector.
25, the data acquisition unit 491 randomly sets sampling points that serve as starting points of time-series data within the coordinate space indicated by the state change speed vector. The data acquisition unit 491 sets the randomly set sampling points as the starting point. As in the case of FIG. 24, the data acquisition unit 491 repeatedly sets new sampling points and directed edges by moving from the starting point along the state change speed vector.
For example, the data acquisition unit 491 repeats the setting of random sampling points a predetermined number of times to acquire a graph based on a predetermined number of pieces of time-series data.

As described above, the data acquisition unit 491 acquires a first graph in which the time series data f ^m ts _i to be predicted is shown in a first coordinate space. The graph generation unit 492 performs coordinate conversion from the first coordinates to the second coordinates using a distance index that is a weighted sum of the distance in the first coordinate space between two pieces of data included in the time series data ^{f m ts} _i shown in the first coordinate space and the time interval in the time series data f m ts _i between those two pieces of data, such that the smaller the value of the distance index, the closer the distance between the two pieces of data in the second coordinate space, thereby generating a second graph in which the time series data f ^* ts _i is shown in the second coordinate space.

The graph generating device 400 can visualize the predicted state changes and show the relative length of time it takes for the state changes to occur.
Specifically, the graph generation device 400 generates a state change graph that shows how the state of the graph generation target changes and includes information indicating the state, such as the name of the state. A person or a state prediction device can plot the current state of the prediction target on the state change graph to predict the future state.
In this case, since the state change graph is generated so that the distance in the coordinate space of the state change graph is proportional to the time interval in the input time series data, it is expected that a person or a state prediction device will be able to predict the time required for a state change with relatively high accuracy.

In addition, the data acquisition unit 491 acquires tidal flow chart data, which is data indicating state change rate vectors at each point within the first coordinate space, and plots state time series data within the first coordinate space based on the state change rate vectors to generate a first graph.
The graph generating unit 492 can convert the first graph acquired by the data acquiring unit 491 to generate a state change graph in which the relative lengths of time taken for state changes are shown in the diagram.

Figure 26 is a schematic block diagram illustrating an example of a computer configuration according to at least one embodiment. In the configuration shown in Figure 26, the computer 700 includes a CPU 710, a main memory device 720, an auxiliary memory device 730, and an interface 740.

Any one or more of the graph generation device 100, graph generation device 100b, state prediction device 300, and graph generation device 400 described above may be implemented in the computer 700. In this case, the operation of each of the processing units described above is stored in the auxiliary storage device 730 in the form of a program. The CPU 710 reads the program from the auxiliary storage device 730, expands it in the main storage device 720, and executes the above processing in accordance with the program. The CPU 710 also allocates storage areas in the main storage device 720 corresponding to each of the storage units described above in accordance with the program.

When the graph generation device 100 is implemented in a computer 700, the operation of the control unit 190 and each of its components is stored in the auxiliary storage device 730 in the form of a program. The CPU 710 reads the program from the auxiliary storage device 730, loads it into the main storage device 720, and executes the above-mentioned processing in accordance with the program.

Furthermore, the CPU 710 allocates a storage area corresponding to the storage unit 180 in the main storage device 720 in accordance with the program.
Communication with other devices via the communication unit 110 is performed by the interface 740 having a communication function and performing communication under the control of the CPU 710 .
The display on the display unit 120 is implemented by the interface 740 having a display device and displaying various images under the control of the CPU 710 .
The operation input unit 130 receives a user operation by the interface 740 having input devices such as a keyboard and a mouse, and outputs information indicating the received user operation to the CPU 710 .

When the graph generation device 100b is implemented in a computer 700, the operation of the control unit 190b and each of its components is stored in the auxiliary storage device 730 in the form of a program. The CPU 710 reads the program from the auxiliary storage device 730, loads it into the main storage device 720, and executes the above-mentioned processing in accordance with the program.

Furthermore, the CPU 710 allocates a storage area corresponding to the storage unit 180 in the main storage device 720 in accordance with the program.
Communication with other devices via the communication unit 110 is performed by the interface 740 having a communication function and performing communication under the control of the CPU 710 .
The display on the display unit 120 is implemented by the interface 740 having a display device and displaying various images under the control of the CPU 710 .
The operation input unit 130 receives a user operation by the interface 740 having input devices such as a keyboard and a mouse, and outputs information indicating the received user operation to the CPU 710 .

When the state prediction device 300 is implemented in a computer 700, the operation of the control unit 390 and each of its components is stored in the auxiliary storage device 730 in the form of a program. The CPU 710 reads the program from the auxiliary storage device 730, loads it into the main storage device 720, and executes the above-mentioned processing in accordance with the program.

Furthermore, the CPU 710 allocates a storage area corresponding to the storage unit 380 in the main storage device 720 in accordance with the program.
Communication with other devices by the communication unit 310 is performed by the interface 740 having a communication function and performing communication under the control of the CPU 710 .
The display unit 320 displays various images by having an interface 740 with a display device and displaying various images under the control of the CPU 710 .
The operation input unit 330 receives a user operation by the interface 740 having input devices such as a keyboard and a mouse, and outputs information indicating the received user operation to the CPU 710 .

When the graph generation device 400 is implemented in a computer 700, the operation of the control unit 490 and each of its components is stored in the auxiliary storage device 730 in the form of a program. The CPU 710 reads the program from the auxiliary storage device 730, loads it into the main storage device 720, and executes the above-mentioned processing in accordance with the program.

Furthermore, the CPU 710 allocates a storage area corresponding to the storage unit 480 in the main storage device 720 in accordance with the program.
Communication with other devices by the communication unit 410 is performed by the interface 740 having a communication function and performing communication under the control of the CPU 710 .
The display unit 420 displays various images by having the interface 740 have a display device and displaying various images under the control of the CPU 710 .
The operation input unit 430 receives a user operation by the interface 740 having input devices such as a keyboard and a mouse, and outputs information indicating the received user operation to the CPU 710 .

Note that a program for realizing all or part of the functions of the graph generation device 100, the graph generation device 100b, the state prediction device 300, and the graph generation device 400 may be recorded on a computer-readable recording medium, and the program recorded on this recording medium may be read into a computer system and executed to perform processing of each part. Note that the term "computer system" here includes hardware such as an OS (Operating System) and peripheral devices.
Furthermore, "computer-readable recording medium" refers to portable media such as flexible disks, optical magnetic disks, ROMs (Read Only Memory), and CD-ROMs (Compact Disc Read Only Memory), as well as storage devices such as hard disks built into computer systems. The program may be one that realizes part of the functions described above, or may be one that can realize the functions described above in combination with a program already recorded in the computer system.

The above describes an embodiment of the present invention in detail with reference to the drawings, but the specific configuration is not limited to this embodiment and includes design modifications and the like that do not deviate from the gist of the present invention.

1 State prediction system 100, 100b, 400 Graph generation device 110, 310, 410 Communication unit 120, 320, 420 Display unit 130, 330, 430 Operation input unit 180, 380, 480 Memory unit 190, 190b, 390, 490 Control unit 191, 391, 491 Data acquisition unit 192, 492 Graph generation unit 193, 493 State information addition unit 194 Learning control unit 210 Feature extraction unit 210a Autoencoder 210b GCM
211a, 211b Encoder 212a, 212b Relay layer 213a, 213b Decoder 214b Multiplier 215b Average calculation unit 216b Argmax calculation unit 220 Graph formation unit 300 State prediction device 392 Prediction unit

Claims (10)

a data acquisition unit that acquires time series data related to a prediction target;
a graph generating unit that converts each of the data included in the time series into a feature using a feature extraction method obtained by learning using an evaluation index in which an evaluation is higher the closer the distance in a feature space between two feature amounts obtained by extracting feature amounts from each of the two data included in the same time series is to a value obtained by multiplying a time interval in the time series of the two data by a predetermined proportionality coefficient ; and
a state information adding unit that adds, to a portion of the coordinate space, information indicating a state of the prediction target at a time in the time series of the data that is converted into a coordinate value included in the portion;
A graph generating device comprising: a data acquisition unit that acquires a first graph in which a time series of data related to a prediction target is shown in a first coordinate space ;
a graph generating unit that generates a second graph in which the time series is displayed in the second coordinate space by performing coordinate transformation from the first coordinates to second coordinates using a distance index that is a weighted sum of a distance in the first coordinate space between two pieces of data included in the time series displayed in the first coordinate space and a time interval in the time series between the two pieces of data, such that the smaller the value of the distance index, the closer the distance between the two pieces of data in the second coordinate space;
a state information adding unit that adds, to a portion of the coordinate space, information indicating a state of the prediction target at a time in the time series of the data that is converted into a coordinate value included in the portion;
A graph generating device comprising : the data acquisition unit acquires tidal flow chart data that is data indicating a state change speed vector at each point in the first coordinate space, and plots time-series data of the state in the first coordinate space based on the state change speed vector to generate the first graph;
The graph generating device according to claim 2 . a prediction unit that performs a state prediction on a time series of data related to a prediction target using a graph whose coordinate space is the feature space obtained by converting each of the data included in the time series into feature quantities using a feature extraction method obtained by learning using an evaluation index in which an evaluation is higher the closer the distance in a feature space between two feature quantities obtained by extracting feature quantities from each of two data included in the same time series among the time series of data related to the prediction target is, the closer the distance in a feature space between the two feature quantities is to a value obtained by multiplying a predetermined proportionality coefficient in the time series of the two data. 5. The state prediction device according to claim 4, wherein the prediction unit detects the data at a smallest distance in the coordinate space from a point at which the state of the prediction target before a state change indicated as the state prediction target is plotted in the coordinate space , and predicts the state of the prediction target after the state change based on an edge in the graph from the detected data to another data. 5. The state prediction device according to claim 4, wherein the prediction unit detects the data located within a predetermined distance in the coordinate space from a point at which a state of the prediction target before a state change indicated as a state prediction target is plotted in the coordinate space , and predicts the state of the prediction target after a state change based on an edge of the graph from the detected data to another data. The computer
Obtaining a time series of data relating to a target for prediction;
converting each of the data included in the time series into the feature using a feature extraction method obtained by learning using an evaluation index in which the closer the distance in a feature space between two feature amounts obtained by performing feature extraction on each of the two data included in the same time series is to a value obtained by multiplying a predetermined proportionality coefficient by the time interval in the time series of the two data, the higher the evaluation ; and generating a graph in which the feature space is a coordinate space;
adding information to a portion of the coordinate space that indicates a state of the prediction target at a time in the time series of the data that is converted into a coordinate value included in the portion;
A graph generation method including: The computer
a feature extraction method obtained by learning using an evaluation index in which an evaluation is higher the closer the distance in a feature space between two feature values obtained by performing feature extraction on each of two data included in the same time series of data related to a prediction target is to a value obtained by multiplying a time interval in the time series of the two data by a predetermined proportional coefficient, and converting each of the data included in the time series into the feature values, and predicting a state of the prediction target using a graph in which the feature space is a coordinate space . On the computer,
Obtaining a time series of data relating to a target for prediction;
converting each of the data included in the time series into the feature using a feature extraction method obtained by learning using an evaluation index in which the closer the distance in a feature space between two feature amounts obtained by performing feature extraction on each of the two data included in the same time series is to a value obtained by multiplying a predetermined proportionality coefficient by the time interval in the time series of the two data, the higher the evaluation ; and generating a graph in which the feature space is a coordinate space;
adding information to a portion of the coordinate space that indicates a state of the prediction target at a time in the time series of the data that is converted into a coordinate value included in the portion;
A program to execute. On the computer,
A program for executing the following: predicting a state of a prediction target using a graph whose coordinate space is the feature space obtained by converting each of the data included in the time series into a feature using a feature extraction method obtained by learning using an evaluation index in which the evaluation is higher the closer the distance in feature space between two feature values obtained by extracting features from each of two data included in the same time series of data related to the prediction target is to a value obtained by multiplying the time interval in the time series of the two data by a predetermined proportional coefficient .