JP7412283B2 - Prediction model generation system and prediction system - Google Patents

Description

The present invention relates to a predictive model generation system and a predictive system.

A system for predicting power demand, which is one of the statistical values to be predicted, is known. For example, Patent Document 1 discloses a system that predicts power demand based on factors that are explanatory variables. This system attempts to reduce prediction errors by appropriately evaluating the amount of variation in factors.

JP2020-64446A

By the way, movement is one of the basic actions in people's social lives, and movement of people affects various phenomena related to economic activities, transportation, public health, etc. On the other hand, it is desired to accurately predict various statistical values related to people's social lives.

SUMMARY OF THE INVENTION The present invention has been made in view of the above-mentioned problems, and it is an object of the present invention to accurately predict statistical values regarding people's social lives by using information regarding people's movements.

In order to solve the above problems, a predictive model generation system according to one embodiment of the present invention is a predictive model generation system that generates a predictive model for predicting an activity statistical value that is a statistical value associated with a location, Among the performance statistics that are the performance statistics of past activity statistics before the reference time, the performance statistics of other locations other than one of the multiple locations are unified based on the inter-location relationship information. a first feature data generation unit that generates feature data representing the characteristics of the activity statistics of one location by reflecting the performance statistics of the locations; Prediction that generates a predictive model using learning data consisting of a first feature data generation unit that is information indicating the number of people moving, feature data of each location, and correct data consisting of performance statistics at a reference time for each location. The present invention includes a predictive model generating unit that updates parameters of the predictive model based on an error between a model output value obtained by inputting feature data into the predictive model and correct data.

According to the above embodiment, the characteristic data is generated by reflecting the performance statistics of one place on the performance statistics of another place according to the number of people moving between the places. The feature data generated in this way constitutes a feature amount indicating the characteristics of the activity statistics of one place, and this feature amount includes the influence of the activity statistics of other places. Then, the predictive model is developed by performing machine learning based on the error between the model output value obtained by inputting the feature data of each location into the predictive model and the correct data consisting of the actual statistical values at the reference time of each location. Constructed. Therefore, it is possible to obtain a prediction model that can accurately predict activity statistics associated with a location.

By using information about people's movements, it becomes possible to predict statistical values regarding people's social lives with high accuracy.

FIG. 1 is a block diagram showing the functional configuration of a predictive model generation device according to the present embodiment. It is a block diagram showing the functional composition of the prediction device of this embodiment. It is a hardware block diagram of a prediction model generation device and a prediction device. It is a figure showing the structure of stay information stored in a movement information storage part, and an example of data included. FIG. 3 is a diagram illustrating an example of a table of the number of people traveling between cities generated based on stay information. FIG. 2 is a diagram illustrating an example of a city graph in which cities are nodes and the number of people moving between cities is an edge. FIG. 6A is a diagram showing an example of a directed city graph in which the direction of movement of people between cities is taken into consideration. FIG. 6(b) is a diagram illustrating an example of an undirected city graph in which only the amount of movement of people between cities is considered, without considering the direction of movement of people between cities. FIG. 3 is a diagram illustrating a first example of an outline of generation of feature data and generation of a predictive model. FIG. 3 is a diagram illustrating an example of input feature amounts input to the neural network in the learning phase of the prediction model of the first example. FIG. 6 is a diagram illustrating an example of input feature amounts that are input to a learned prediction model in the phase of predicting activity statistics in the first example. FIG. 7 is a diagram illustrating a second example of an outline of generation of feature data and generation of a predictive model. It is a figure which shows the example of the input feature-value input into a convolutional neural network and a neural network in a 2nd example. FIG. 2 is a diagram for explaining an outline of a convolutional neural network (graph convolutional neural network) for a graph structure. FIG. 3 is a diagram illustrating a convolutional neural network algorithm for a graph structure. 3 is a flowchart illustrating feature data generation processing using a convolutional neural network for a graph structure. 3 is a flowchart illustrating learning processing of a convolutional neural network and a neural network of a predictive model for a graph structure. FIG. 7 is a diagram illustrating an example of input feature amounts indicating the characteristics of each city in the second example of predicting activity statistics. It is a flowchart which shows the processing content of the predictive model generation method in a predictive model generation device. It is a flowchart which shows the processing content of the prediction method in a prediction device. FIG. 19(a) is a diagram showing the configuration of the predictive model generation program. FIG. 19(b) is a diagram showing the configuration of the prediction program.

Embodiments of a predictive model generation system and a predictive system according to the present invention will be described with reference to the drawings. In addition, if possible, the same parts are given the same reference numerals and redundant explanations will be omitted.

FIG. 1 is a diagram showing the device configuration of a predictive model generation system and the functional configuration of the predictive model generation device according to the present embodiment. As shown in FIG. 1, the predictive model generation system 1 includes a predictive model generating device 10, a movement information storage section 31, an inter-location relationship information storage section 32, a performance statistical value information storage section 33, and a predictive model storage section 34. . The predictive model generation device 10 may constitute a predictive model generation system 1.

The predictive model generation device 10 is a device that generates a predictive model for predicting activity statistics, which are statistical values associated with locations. The activity statistical value is a statistical value associated with a place, and is a statistical value regarding the social life, activities, etc. of people in the place. The activity statistics include, for example, economic index values and the number of disease patients. In this embodiment, a prediction model that predicts activity statistics of a city as an example of a location is generated.

The predictive model generation device 10 functionally includes a movement information acquisition section 11, an inter-location relationship information generation section 12, a first performance statistical value acquisition section 13, a first feature data generation section 14, a prediction model generation section 15, and A prediction model output unit 16 is provided. Each of the functional units 11 to 16 of the predictive model generation device 10 is configured to be able to access storage devices such as a movement information storage unit 31, an inter-location relationship information storage unit 32, an actual statistical value information storage unit 33, and a predictive model storage unit 34. has been done. Each of the functional units 11 to 16 of the predictive model generation device 10 may be configured in one device, or may be distributed and configured in a plurality of devices.

FIG. 2 is a diagram showing the device configuration of the prediction system and the functional configuration of the prediction device according to the present embodiment. As shown in FIG. 2, the prediction system 2 includes a prediction device 20, an inter-location relationship information storage section 32, a performance statistical value information storage section 33, and a prediction model storage section 34. The prediction device 20 may constitute the prediction system 2.

The prediction device 20 is a device that predicts activity statistics using a prediction model that has been learned by machine learning. The prediction device 20 functionally includes a second performance statistical value acquisition section 21, a second feature data generation section 22, a prediction section 23, and an output section 24.

Each of the functional units 21 to 24 of the prediction device 20 is configured to be able to access storage devices such as an inter-location relationship information storage unit 32, an actual statistical value information storage unit 33, and a prediction model storage unit . Each of the functional units 21 to 24 of the prediction device 20 may be configured in one device, or may be distributed and configured in a plurality of devices.

Note that the block diagrams shown in FIGS. 1 and 2 show blocks in functional units. These functional blocks (components) are realized by any combination of at least one of hardware and software. Furthermore, the method for realizing each functional block is not particularly limited. That is, each functional block may be realized using one physically or logically coupled device, or may be realized using two or more physically or logically separated devices directly or indirectly (e.g. , wired, wireless, etc.) and may be realized using a plurality of these devices. The functional block may be realized by combining software with the one device or the plurality of devices.

Functions include judgment, decision, judgment, calculation, calculation, processing, derivation, investigation, exploration, confirmation, reception, transmission, output, access, resolution, selection, selection, establishment, comparison, assumption, expectation, consideration, These include, but are not limited to, broadcasting, notifying, communicating, forwarding, configuring, reconfiguring, allocating, mapping, and assigning. I can't do it. For example, a functional block (configuration unit) that performs transmission is called a transmitting unit or a transmitter. In either case, as described above, the implementation method is not particularly limited.

For example, the predictive model generation device 10 and the predictive device 20 in one embodiment of the present invention may function as a computer. FIG. 3 is a diagram showing an example of the hardware configuration of the predictive model generation device 10 and the prediction device 20 according to the present embodiment. The predictive model generation device 10 and the predictive device 20 may be physically configured as a computer device including a processor 1001, a memory 1002, a storage 1003, a communication device 1004, an input device 1005, an output device 1006, a bus 1007, etc. .

In addition, in the following description, the word "apparatus" can be read as a circuit, a device, a unit, etc. The hardware configurations of the prediction model generation device 10 and the prediction device 20 may be configured to include one or more of the devices shown in FIGS. 1 and 2, or may be configured to include some of the devices. You can.

Each function in the prediction model generation device 10 and the prediction device 20 is performed by loading predetermined software (programs) onto hardware such as a processor 1001 and a memory 1002, so that the processor 1001 performs calculations and performs communication and communication by the communication device 1004. This is realized by controlling reading and/or writing of data in the memory 1002 and storage 1003.

The processor 1001, for example, operates an operating system to control the entire computer. The processor 1001 may be configured with a central processing unit (CPU) that includes interfaces with peripheral devices, a control device, an arithmetic device, registers, and the like. For example, each of the functional units 11 to 16, 21 to 24, etc. shown in FIGS. 1 and 2 may be implemented by the processor 1001.

Further, the processor 1001 reads programs (program codes), software modules, and data from the storage 1003 and/or the communication device 1004 to the memory 1002, and executes various processes in accordance with these. As the program, a program that causes a computer to execute at least part of the operations described in the above embodiments is used. For example, each of the functional units 11 to 16 and 21 to 24 of the predictive model generation device 10 and the predictive device 20 may be realized by a control program stored in the memory 1002 and operated by the processor 1001. Although the various processes described above have been described as being executed by one processor 1001, they may be executed by two or more processors 1001 simultaneously or sequentially. Processor 1001 may be implemented with one or more chips. Note that the program may be transmitted from a network via a telecommunications line.

The memory 1002 is a computer-readable recording medium, and includes at least one of ROM (Read Only Memory), EPROM (Erasable Programmable ROM), EEPROM (Electrically Erasable Programmable ROM), RAM (Random Access Memory), etc. may be done. Memory 1002 may be called a register, cache, main memory, or the like. The memory 1002 can store executable programs (program codes), software modules, and the like to implement the predictive model generation method and prediction method according to an embodiment of the present invention.

The storage 1003 is a computer-readable recording medium, such as an optical disk such as a CD-ROM (Compact Disc ROM), a hard disk drive, a flexible disk, a magneto-optical disk (such as a compact disk, a digital versatile disk, or a Blu-ray disk). (registered trademark) disk), smart card, flash memory (eg, card, stick, key drive), floppy disk, magnetic strip, etc. Storage 1003 may also be called an auxiliary storage device. The storage medium mentioned above may be, for example, a database including memory 1002 and/or storage 1003, a server, or other suitable medium.

The communication device 1004 is hardware (transmission/reception device) for communicating between computers via a wired and/or wireless network, and is also referred to as, for example, a network device, network controller, network card, communication module, or the like.

The input device 1005 is an input device (eg, keyboard, mouse, microphone, switch, button, sensor, etc.) that accepts input from the outside. The output device 1006 is an output device (for example, a display, a speaker, an LED lamp, etc.) that performs output to the outside. Note that the input device 1005 and the output device 1006 may have an integrated configuration (for example, a touch panel).

Further, each device such as the processor 1001 and the memory 1002 is connected by a bus 1007 for communicating information. The bus 1007 may be configured as a single bus or may be configured as different buses between devices.

Further, the prediction model generation device 10 and the prediction device 20 include a microprocessor, a digital signal processor (DSP), an application specific integrated circuit (ASIC), a programmable logic device (PLD), a field programmable gate array (FPGA), etc. The hardware may be configured to include hardware, and a part or all of each functional block may be realized by the hardware. For example, processor 1001 may be implemented with at least one of these hardware.

Next, each functional unit of the prediction model generation device 10 and the prediction device 20 will be explained. The movement information acquisition unit 11 acquires stay information indicating that a person has stayed at each location. In this embodiment, the movement information acquisition unit 11 acquires stay information regarding a city as an example of a location. Specifically, the movement information acquisition unit 11 acquires stay information from the movement information storage unit 31. The movement information storage unit 31 is a storage means that stores stay information.

FIG. 4 is a diagram showing an example of stay information stored in the movement information storage section 31. As shown in FIG. 4, the stay information includes a UID (user ID) that identifies the user and a place (city) where the user stayed for each time stamp indicating the date and time. The example shown in FIG. 4 shows that the user identified by UID "001" stayed in the city "A city" on "2020/4/1 12:00". Furthermore, the user with UID "001" stayed in the city "B city" on "2020/4/1 15:00" and stayed in the city "C city" on "2020/4/1 18:00". It is shown. That is, the stay information shown in FIG. 4 indicates that the user with UID "001" moved from City A to City B, and then further moved to City C. Therefore, stay information has the aspect of information indicating the movement of a person.

The inter-place relationship information generation unit 12 generates inter-place relationship information indicating the number of people moving between places based on the stay information. In the present embodiment, the inter-location relationship information generation unit 12 generates a number-of-travel table indicating the number of people moving between cities as a form of inter-location relationship information. The inter-location relationship information generation unit 12 may store the generated moving number table in the inter-location relationship information storage unit 32. The inter-place relationship information storage unit 32 is a storage means for storing inter-place relationship information.

Specifically, the inter-location relationship information generation unit 12 refers to the stay information and counts the number of users for each combination of a source city and a destination city. For example, in the example shown in FIG. 4, the location of the user with UID "001" transitions from "City A" to "City B" according to the time series shown in the time stamp, so the relationship between locations is The information generation unit 12 counts the users with UID "001" as the number of people who moved from "City A" to "City B".

FIG. 5 is a diagram showing an example of a table of the number of people traveling between cities generated based on stay information. As shown in FIG. 5, the total number of counted users is associated with each combination of a source city and a destination city. For example, the number of people who moved from "City A" to "City B" is "100."

The inter-location relationship information generation unit 12 generates a city graph with cities as nodes and the number of people moving between cities as edges, based on the table of people moving. The city graph constitutes one aspect of inter-place relationship information. The inter-place relationship information generation unit 12 may store the generated city graph in the inter-place relationship information storage unit 32. FIG. 6 is a diagram showing an example of a city graph.

FIG. 6A is a diagram showing an example of a directed city graph in which the direction of movement of people between cities is taken into consideration. As shown in FIG. 6(a), the inter-location relationship information generation unit 12 sets A city, B city, and C city as nodes ca, cb, and cc, respectively, and creates an edge between cities where people have moved. to generate a city graph crA. As shown in FIG. 6A, edge e11 indicates that the number of people who moved from city B to city A is 300. Edge e12 indicates that the number of people who moved from city A to city B is 100. Edge e21 indicates that the number of people who moved from city B to city C is 30. Edge e22 indicates that the number of people who moved from City C to City B is 20. Edge e31 indicates that the number of people who moved from city A to city C is 50. Edge e32 indicates that the number of people who moved from City C to City A is 150.

Since the city graph crA is a directed city graph that takes into account the movement direction of people between cities, for example, as shown in edge e11 and edge e12, the number of people who moved from city B to city A and A The number of people who moved from city to city B is treated as a different edge.

FIG. 6(b) is a diagram illustrating an example of an undirected city graph in which only the amount of movement of people between cities is considered, without considering the direction of movement of people between cities. As shown in FIG. 6(b), the inter-location relationship information generation unit 12 sets A city, B city, and C city as nodes ca, cb, and cc, respectively, and creates an edge between cities where people have moved. to generate a city graph crB. As shown in FIG. 6(b), edge e1 indicates that the number of people who moved between City A and City B is 400. Edge e2 indicates that the number of people who moved between City B and City C is 50. Edge e3 indicates that the number of people who moved between City A and City C is 200.

Since the city graph crB is an undirected city graph in which the direction of movement of people between cities is not considered, the edge e1 is the number of people who moved from city B to city A and the number of people who moved from city A to city B. It shows the total number of people moving.

Referring again to FIG. 1, the first performance statistics acquisition unit 13 acquires performance statistics that are performance statistics of activity statistics in the past before the reference time. The first feature data generation unit 14 converts the performance statistics of other places other than one of the plurality of places, among the performance statistics, into the performance statistics of one place based on the inter-place relationship information. By reflecting this in the value, characteristic data representing the characteristics of the activity statistics of one place is generated. The prediction model generation unit 15 generates a prediction model using learning data consisting of a pair of feature data of each location and correct data consisting of performance statistics at a reference time of each location.

A first example of acquisition of performance statistics, generation of feature data, and generation of a prediction model will be described with reference to FIGS. 7 to 9. FIG. 7 is a diagram schematically showing generation of feature data and prediction model generation in the first example. FIG. 8 is a diagram illustrating an example of input feature amounts input to the neural network in the learning phase of the prediction model in the first example.

The first performance statistic value acquisition unit 13 acquires a performance statistic value as1 that is a performance value of activity statistics in the past before the reference time. The performance statistics information storage unit 33 shown in FIGS. 1 and 2 is a storage means that stores performance statistics. The first performance statistical value acquisition unit 13 may acquire performance statistical values from the performance statistical value information storage unit 33.

If the activity statistical value to be predicted is a statistical value whose change tendency repeats every certain unit period, it is used as an input feature in the learning of the prediction model and the prediction of the activity statistical value. The performance statistical value may include an element of a trend component and an element of a seasonal component.

Specifically, the first performance statistical value acquisition unit 13 calculates the period corresponding to the period to which the reference time belongs in the current period in the previous period which is the unit period before the current period which is the unit period to which the reference time belongs. The actual value of the activity statistics may be acquired as the actual statistical value corresponding to the seasonal component. Further, the first performance statistic value acquisition unit 13 may acquire performance values of time-series activity statistics for a predetermined period in the past before the reference time as performance statistics corresponding to the trend component.

A time series model called a so-called seasonal adjustment model is a model for a single series, but in this embodiment, by associating time series models of multiple locations using inter-location relationship information, the time series model is a seasonal adjustment model. A similar concept of processing is extended to multiple series (locations, cities).

In this embodiment, as shown in FIG. 8, the first performance statistical value acquisition unit 13 obtains the performance statistics of the activity statistics for the same month of the previous year at the reference time and the time series of each month from 1 to N months before the reference time. The actual value of the activity statistics of is obtained for each city as the actual statistical value as1.

The first feature data generation unit 14 classifies each location into a plurality of groups based on the inter-location relationship information, and associates a group identifier for identifying the group with each location. In the present embodiment, the first feature data generation unit 14 performs clustering cl on cities based on the inter-location relationship information constituted by the city graph cr1, and assigns cluster IDs (gid) for identifying clusters to each city. Associate with a city. The city graph cr1 may be the directed city graph crA illustrated in FIG. 6(a), or may be the undirected city graph crB illustrated in FIG. 6(b).

The clustering cl is performed using a predetermined clustering method using the number of people moving between cities, which is represented as an edge in the city graph cr1, as a feature of each city. Although the clustering method applied here is not limited, for example, it may be a clustering method based on an index value called modularity (Q), and this method includes the first method (MEJ Newman: Finding community structure in networks using the eigenvectors of matrices, Phys. Rev. E 74, 036104 (2006)), or the second method (CNM algorithm: A Clauset, MEJ Newman and C Moore: Finding structure community in very large networks. Phys Rev E 70, 066111 (2004)) may be used.

式（１）において、ｅ_ｉｉの項は、総エッジ数に対するクラスタ内エッジ数の割合を表し、ｅ_ｉｊの項は、総エッジ数に対するクラスタ間エッジ数の割合を表す。本実施形態のクラスタリング手法の一例では、クラスタの区切り方を変更してモジュラリティＱを計算し、モジュラリティＱが最大になる都市の組み合わせを逐次探索する。この逐次探索は、上記第１の手法及び第２の手法により効率化される。 In clustering nodes represented by a graph, it is preferable that the edge density within a cluster is high and the edge density between clusters is low. Modularity Q is a value that quantifies the desirability regarding such edge density. Modularity Q is calculated, for example, using the following equation (1).

In Equation (1), the term e _ii represents the ratio of the number of intra-cluster edges to the total number of edges, and the term e _ij represents the ratio of the number of inter-cluster edges to the total number of edges. In an example of the clustering method of this embodiment, the modularity Q is calculated by changing the way the clusters are divided, and the combination of cities with the maximum modularity Q is sequentially searched for. This sequential search is made more efficient by the first method and the second method.

In clustering cl, each city is classified into one of the communities (clusters), and a cluster ID (gid) is given to the cities included in each community. The first feature data generation unit 14 associates the performance statistical value as1 of each city with the cluster ID (gid) assigned to the city, and generates feature data fv1 representing the characteristics of each city. As shown in FIG. 8, the first feature data generation unit 14 associates the cluster ID (gid) assigned to each city with the performance statistical value as1 of each city, along with the place ID that identifies each city. Then, feature data fv1 is generated.

Further, in the input feature amount shown in FIG. 8, correct data tv1 is associated with feature data fv1. The correct data tv1 is data consisting of actual statistical values of each city at the reference time. The feature data fv1 and the correct answer data tv1 constitute learning data used for learning the neural network. Note that the input feature amount including the feature data fv1 and the correct answer data tv1 is data expressed by a vector that can be input to the neural network.

The predictive model generation unit 15 inputs the feature data fv1 to the neural network nn1 included in the predictive model, and obtains a model output value mv1 that is an output from the neural network nn1. Although the type of neural network nn1 applied in this embodiment is not limited, it may be a neural network configured by, for example, a multilayer perceptron. Note that the prediction model may be a machine learning model constructed based on a decision tree. Machine learning models built using decision trees are, for example, XGBoost (Chen, Tianqi, and Carlos Guestrin. "Xgboost: A scalable tree boosting system." Proceedings of the 22nd ACM SIGKDD international conference on knowledge discovery and data mining. 2016. ) and LightGBM (Ke, Guolin, et al. "Lightgbm: A highly efficient gradient boosting decision tree." Advances in neural information processing systems. 2017.).

The predictive model generation unit 15 performs error calculation bp1 between the model output value mv1 and the correct data tv1, and updates the parameters of the neural network nn1 by using the error backpropagation method based on the obtained error. Learn.

The prediction model output unit 16 causes the prediction model storage unit 34 to store the prediction model including the neural network nn1 learned by the prediction model generation unit 15. The predictive model storage unit 34 is a storage unit for storing learned predictive models.

Next, the aspect of predicting activity statistics using the learned prediction model of the first example will be described. The second performance statistic value acquisition unit 21 acquires the performance statistic value as1, which is the performance value of the activity statistics in the past before the prediction target time, from the performance statistic information storage unit 33. Specifically, similar to the learning phase of the prediction model, the second performance statistic value acquisition unit 21 acquires the actual activity statistics of the same month of the previous year at the time of the prediction target and 1 to N months before the prediction target time. Acquire the actual value of activity statistics for each month in time series for each city to be predicted.

The second feature data generation unit 22 calculates the performance statistics of a place other than one of the plurality of places, among the past performance statistics before the prediction target time, based on the inter-place relationship information. By reflecting this on the performance statistics of one place, characteristic data representing the characteristics of the activity statistics of one place is generated.

Specifically, in the aspect of predicting activity statistics in the first example, the second feature data generation unit 22 uses the performance statistics as1 of the city to be predicted acquired by the second performance statistics acquisition unit 21. By associating the cluster ID (gid) given to the city, feature data fv1 representing the characteristics of the city to be predicted is generated. By associating the cluster ID (gid) with the actual statistical value as1, the feature data of other cities having the same cluster ID (gid) as the city to be predicted is added to the feature data of the city to be predicted. ). Therefore, in predicting activity statistics, the characteristic data of other cities is appropriately reflected in the characteristic data of the city to be predicted.

FIG. 9 is a diagram illustrating an example of input feature amounts input to a learned prediction model in the phase of predicting activity statistics in the first example. As shown in FIG. 9, the second feature data generation unit 22 associates the performance statistical value as1 of each city with the cluster ID (gid) of the city, and generates feature data fv1 representing the characteristics of each city. .

The prediction unit 23 inputs the feature data fv1 generated by the second feature data generation unit 22 into a prediction model to obtain a prediction statistical value. Specifically, the prediction unit 23 inputs the feature data fv1 into the neural network nn1 of the prediction model stored in the prediction model storage unit 34, and acquires the prediction statistical value output from the neural network nn1.

The output unit 24 outputs the predicted statistical value acquired by the prediction unit 23 in a predetermined manner as a predicted result of the activity statistical value. Specifically, the output unit 24 outputs the predicted results of the activity statistics by displaying the predicted statistics on a predetermined display device or storing the predicted statistics in a predetermined storage means.

Next, a second example of acquisition of performance statistics, generation of feature data, and generation of a prediction model will be described with reference to FIGS. 10 to 16. FIG. 10 is a diagram schematically showing generation of feature data and prediction model generation in the second example. FIG. 11 is a diagram illustrating an example of generation of feature data and input feature amounts provided to the neural network in the learning phase of the predictive model in the second example.

The first performance statistic value acquisition unit 13 acquires a performance statistic value as2 that is a performance value of activity statistics in the past before the reference time. Specifically, the first performance statistical value acquisition unit 13 acquires the performance statistical value as2 from the performance statistical value information storage unit 33. As shown in FIG. 11, the first performance statistics acquisition unit 13 acquires the performance statistics of the activity statistics of the same month of the previous year at the reference time and the activity statistics of each month in the time series from 1 to N months before the reference time. A track record value is acquired for each city as a track record statistical value as2. In addition, as shown in FIG. 11, the first performance statistics acquisition unit 13 acquires the performance statistics of each city at the reference time as correct data tv2, and stores the results of each city together with the location ID for identifying the city. Correct answer data tv2 is associated with performance statistical value as2 to constitute an input feature quantity.

The first feature data generation unit 14 aggregates performance statistical values of places other than one of the plurality of places using a predetermined method based on inter-place relationship information to generate aggregated statistical values, The feature data fv2 of one place is generated by combining the actual statistical value of one place and the aggregated statistical value using a predetermined method. Specifically, in order to generate the feature data fv2 representing the characteristics of one city, the first feature data generation unit 14 aggregates the performance statistical values as2 of other cities based on the city graph cr2 and generates aggregate statistics. The performance statistical value as2 of one city and the aggregated statistical value are combined to generate characteristic data fv2 of one city. In the second example of this embodiment, the city graph cr2 is an undirected city graph crB as shown in FIG. 6(b), but the city graph cr2 is an undirected city graph crB as shown in FIG. 6(a). It may be a graph crA.

In the second example, the first feature data generation unit 14 includes a graph convolutional neural network (GCN) gcn. GCN is a type of neural network for graph data, which aggregates the features of nodes adjacent to the node of interest using a predetermined aggregation method, and uses the aggregated features as the feature of the node of interest. A convolutional feature of the node of interest is generated by combining. The GCN of this embodiment is applied to a city graph in which each location is a node, the actual statistical value as2 of each location is a node feature, and the number of people moving between locations represented by inter-location relationship information is an edge. .

FIG. 12 is a diagram for explaining an outline of a convolutional neural network (graph convolutional neural network) for a graph structure. FIG. 12 shows a two-layer convolution of a graph gp including nodes na, nb, nc, nd, and ne.

In FIG. 12, when node na is the node of interest, the feature amounts of nodes nb, nc, and nd connected to node na by edges are aggregated using a predetermined aggregation function cn to generate an aggregate feature amount. A convolution feature of the node na is generated by combining the feature of the node na.

In the example shown in FIG. 12, two layers of convolution are performed, so prior to generating the convolution feature of the first layer node na, the convolution feature of each of the nodes nb, nc, and nd is generated. Specifically, the convolutional feature amount of the node nb is generated based on the feature amounts of the nodes na and nc connected to the node nb by an edge. A convolution feature of the node nc is generated based on the features of the nodes na, nb, and ne connected to the node nc by edges. A convolution feature of the node nd is generated based on the feature of the node na connected to the node nd by an edge.

The first feature data generation unit 14 uses the actual statistical value as2 of each location as a node feature amount associated with each node, and connects it to one node corresponding to one location, which is the location of interest, by an edge. By aggregating the node features of other nodes corresponding to other locations based on the inter-location relationship information corresponding to each edge, the aggregated feature of one node corresponds to the aggregated statistical value regarding one location. generate. Next, the first feature data generation unit 14 combines the node feature amount of one node and the aggregated feature amount of one node and multiplies the combined feature amount generated by the weight matrix, thereby generating the number of features of one node. Generate convolutional features. Then, the first feature data generation unit 14 generates the convolutional feature amount of each node as a convolution process for one layer, and converts the generated convolution feature amount of each node into the node feature of each node in the convolution process of the next layer. The convolutional feature amount generated by convolution processing of a predetermined number of layers is defined as feature data fv2 of each location.

A specific example of the convolution feature generation process will be described with reference to FIGS. 13 and 14. FIG. 13 is a diagram showing an algorithm for generating convolutional features using a graph convolutional neural network. The algorithm is made up of lines 1-9.

In line 1, the node feature x _v of node v is assumed to be h ⁰ _v . This node feature corresponds to the node feature of the 0th layer.

In line 2, the number of layers in the hierarchical node feature convolution to be performed is set to K, and while the number of layers is incremented by the variable k, the processes of lines 2 to 8 are repeated for the number of layers. In line 3, v identifies a node, and it is shown that the processing of lines 3 to 6 is repeated for the number of nodes V.

In lines 4 and 5, the first feature data generation unit 14 aggregates the node features of the node of interest (node at one location) and the adjacent node (node at another location) connected by an edge using a predetermined aggregation method. The aggregated feature is then combined with the node feature of the node of interest to generate a convolutional feature of the node of interest.

Specifically, in line 4, the first feature data generation unit 14 generates the node feature amount h ^k−1 _u (from the second layer onward) of all adjacent nodes N(v) adjacent to the node u corresponding to the node of interest. (in the case of convolution, the convolutional features) are aggregated by a predetermined aggregation function AGGREGATE _k to obtain the k-th layer aggregated features h ^k _N(v) . In this embodiment, the predetermined aggregation function AGGREGATE _k may be a weighted average of the node features of adjacent nodes, with the weight being the number of people moving between nodes (cities).

In line 5, the first feature data generation unit 14 combines the node feature amount of the node v of the (k-1)th layer and the aggregated feature amount h ^k _N(v) of the kth layer using a combination function CONCAT. , by the weight matrix W ^k and further applied with the activation function σ to obtain the convolution feature h ^k _v of the node v in the k-th layer. The activation function σ may be, for example, a ReLU function expressed by the following equation (2).

In line 7, the first feature data generation unit 14 normalizes the convolutional feature h ^k _v . Then, in line 9, the first feature data generation unit 14 aggregates the convolutional feature quantities h ^k _v of the node of interest obtained by convolving the node feature quantities of the predetermined K layer using a predetermined aggregation method, It is generated as feature data in a vector representation indicating the features of the node of interest.

FIG. 14 is a diagram illustrating a flowchart of graph convolution feature generation processing. In step S21, the first feature data generation unit 14 acquires a node for which feature data (convolution feature amount) is to be generated.

In step S22, the first feature data generation unit 14 obtains the initial value h ⁰ _v of the feature vector (node feature amount, performance statistical value as2) of the target node. In step S23, the first feature data generation unit 14 obtains adjacent nodes of the target node.

In step S24, the first feature data generation unit 14 acquires the feature vector (node feature amount, performance statistical value as2) of the adjacent node, and performs AGGREGATE processing (weighted average) using the number of people moving between cities as a weight. As a result, the aggregate feature quantity h ^k _N(v) is obtained.

In step S25, the first feature data generation unit 14 concatenates (CONCAT) the aggregated feature amount h ^k _N(v) and the feature vector h ⁰ _v of the target node, multiplies it by the weight matrix W ^k , and further generates an activation function σ to obtain the output h ^k _v .

By repeating the process shown in this flowchart for all nodes, convolution of the node feature values in one layer is realized. Then, by further repeating the processing for a predetermined number of K layers, convolution of the node feature amounts of the K layer of all nodes is realized, and feature data fv2 of each node (city) is obtained.

Note that in the AGGREGATE process in which an aggregated feature is calculated by a weighted average of node features (actual statistical value as2) of adjacent nodes in step S24, the number of people moving between cities is used as a weight. That is, based on the undirected city graph illustrated in FIG. The total number of people moving from the target node to the adjacent node (second moving number) is used as the weight.

On the other hand, based on the directed city graph illustrated in FIG. 6(a), the number of people moving from an adjacent node to a target node and the number of people moving from a target node to an adjacent node are Only the number of people moving from an adjacent node to a target node may be used as a weight for the node feature amount of each adjacent node in the AGGREGATE process regarding the target node.

Referring again to FIG. 10, the predictive model generation unit 15 inputs the feature data fv2 to the neural network nn2 included in the predictive model, and obtains the model output value mv2 that is the output from the neural network nn2. Although the type of neural network nn2 applied in this embodiment is not limited, it may be a neural network composed of, for example, a multilayer perceptron. Note that the prediction model may be a machine learning model constructed based on a decision tree. Machine learning models built using decision trees are, for example, XGBoost (Chen, Tianqi, and Carlos Guestrin. "Xgboost: A scalable tree boosting system." Proceedings of the 22nd ACM SIGKDD international conference on knowledge discovery and data mining. 2016. ) and LightGBM (Ke, Guolin, et al. "Lightgbm: A highly efficient gradient boosting decision tree." Advances in neural information processing systems. 2017.).

The predictive model generation unit 15 performs error calculation bp2 between the model output value mv2 and the correct data tv2 (see FIG. 11), and uses the error backpropagation method based on the obtained error to calculate the parameters of the neural network nn2 and the weight matrix of the GCN. Learning of the neural network nn2 and GCN is performed by updating the parameters of ^Wk .

FIG. 15 is a flowchart showing the learning process of the graph convolutional neural network gcn and the predictive model neural network nn2. In step S31, as described above, the first feature data generation unit 14 generates feature data fv2 (feature vectors) of all nodes (locations (cities)).

上記式（３）におけるかっこ内は、正解データｔｖ２からモデル出力値ｍｖ２を減じた差分を表す。ｎはサンプル数である。 In step S32, the prediction model generation unit 15 inputs the feature data fv2 of each node to the neural network nn2, and calculates a prediction error based on the model output value mv2 output from the neural network nn2 and the correct answer data tv2. The prediction error is calculated by the following equation (3).

The value in parentheses in the above formula (3) represents the difference obtained by subtracting the model output value mv2 from the correct data tv2. n is the number of samples.

In step S33, the predictive model generation unit 15 updates the parameters of the neural network nn2 and the weight matrix ^Wk using the error backpropagation method.

Next, the aspect of predicting activity statistics using the learned prediction model of the second example will be described. The second performance statistic value acquisition unit 21 acquires the performance statistic value as2, which is the performance value of the activity statistics value in the past before the prediction target time, from the performance statistic value information storage unit 33. FIG. 16 is a diagram illustrating an example of input feature amounts indicating the characteristics of each city in the second example of predicting activity statistics. As shown in FIG. 16, the second performance statistics acquisition unit 21 obtains the actual activity statistics of the same month of the previous year at the time of the prediction target, and the activity statistics of each month in the time series from 1 to N months before the prediction target time. The actual value of the value and the actual value are obtained for each city to be predicted. Then, the second performance statistic value acquisition unit 21 associates the acquired performance statistic value as2 with a place ID that identifies the city, and generates an input feature quantity to be input to the GCN.

Similarly to the learning phase of the prediction model, the second feature data generation unit 22 calculates performance statistics of other locations than one of the multiple locations using a predetermined method based on inter-location relationship information. The feature data fv2 of one place is generated by aggregating and generating an aggregate statistical value, and combining the actual statistical value of one place and the aggregate statistical value using a predetermined method. Specifically, the second feature data generation unit 22 aggregates the performance statistical values as2 of other cities based on the city graph cr2 to generate aggregated statistical values, and combines the performance statistical values as2 of the first city with the aggregated statistics. The characteristic data fv2 of one city is generated by combining the values.

More specifically, the second feature data generation unit 22 inputs the input feature amount including the performance statistical value as2 of each city to the learned GCN, and generates the feature data fv2 that is the output from the GCN. The city graph cr2 referred to by the GCN may be generated based on the number of people moving between the corresponding cities at the time of prediction target, or may be generated based on the number of people moving between the corresponding cities at the time of learning the prediction model. It may be generated by

The prediction unit 23 inputs the feature data fv2 generated by the second feature data generation unit 22 into the trained prediction model to obtain prediction statistics. Specifically, the prediction unit 23 inputs the feature data fv2 into the neural network nn2 of the prediction model stored in the prediction model storage unit 34, and acquires the prediction statistical value output from the neural network nn2.

The output unit 24 outputs the predicted statistical value acquired by the prediction unit 23 in a predetermined manner as a predicted result of the activity statistical value. Specifically, the output unit 24 outputs the predicted results of the activity statistics by displaying the predicted statistics on a predetermined display device or storing the predicted statistics in a predetermined storage means.

Note that the prediction model including the neural network nn1 in the first example and the prediction model including the trained GCN and neural network nn2 in the second example are read or referenced by a computer, and the computer performs predetermined processing. It can be understood as a program that is executed and causes a computer to realize a predetermined function.

That is, the learned prediction model and GCN of this embodiment are used in a computer equipped with a CPU and memory. Specifically, the CPU of the computer applies learned weighting coefficients corresponding to each layer to the input features input to the input layer of the neural network according to instructions from a trained predictive model stored in memory. It performs calculations based on response functions, etc., and outputs the results from the output layer.

FIG. 17 is a flowchart showing the processing details of the predictive model generation method in the predictive model generation device 10.

In step S1, the movement information acquisition unit 11 acquires stay information indicating that a person has stayed at each location (city).

In step S2, the inter-place relationship information generation unit 12 generates inter-place relationship information indicating the number of people moving between places (cities) based on the stay information. In this embodiment, the inter-place relationship information generation unit 12 generates a city graph as the inter-place relationship information.

In step S3, the first performance statistic value acquisition unit 13 acquires performance statistic values as1 and as2, which are performance values of activity statistics in the past before the reference time.

In step S4, the first feature data generation unit 14 converts the performance statistics of a place other than one of the plurality of places, among the performance statistics, into one based on the inter-place relationship information. By reflecting the performance statistics of a place, characteristic data fv1 and fv2 representing the characteristics of the activity statistics of one place are generated.

In step S5, the predictive model generation unit 15 performs machine learning using learning data consisting of a pair of feature data fv1, fv2 of each location and correct data tv1, tv2 consisting of performance statistical values at the reference time of each location. A predictive model is generated by

In step S6, the prediction model output unit 16 outputs a prediction model including the neural networks nn1 and nn2 learned by the prediction model generation unit 15. Specifically, the predictive model output unit 16 causes the predictive model storage unit 34 to store the learned predictive model.

FIG. 18 is a flowchart showing the processing details of the prediction method in the prediction device 20. In step S11, the second performance statistics acquisition unit 21 acquires performance statistics as1 and as2, which are performance statistics of activity statistics in the past before the prediction target time, from the performance statistics information storage unit 33 for each city.

In step S12, the second feature data generation unit 22 calculates the performance statistics of a place other than one of the plurality of places among the past performance statistics before the prediction target time based on the relationship between the places. By reflecting the performance statistics of one place based on the sex information, characteristic data fv1 and fv2 representing the characteristics of the activity statistics of one place are generated.

In step S13, the prediction unit 23 inputs the feature data fv1 and fv2 generated by the second feature data generation unit 22 into the prediction model to obtain a prediction statistical value.

In step S14, the output unit 24 outputs the predicted statistical value acquired by the prediction unit 23 in a predetermined manner as a predicted result of the activity statistical value.

Next, a prediction model generation program for causing a computer to function as the prediction model generation device 10 and a prediction program for causing the computer to function as the prediction device 20 of this embodiment will be described. FIG. 19(a) is a diagram showing the configuration of the predictive model generation program. FIG. 19(b) is a diagram showing the configuration of the prediction program.

The predictive model generation program P1 includes a main module m10 that comprehensively controls the predictive model generation process in the predictive model generation device 10, a movement information acquisition module m11, an inter-location relationship information generation module m12, and a first performance statistics acquisition module m13. , a first feature data generation module m14, a predictive model generation module m15, and a predictive model output module m16. Then, each module m11 to m16 generates a movement information acquisition unit 11, an inter-location relationship information generation unit 12, a first performance statistical value acquisition unit 13, a first feature data generation unit 14, a prediction model generation unit 15, and a prediction model output. Each function for section 16 is realized.

The prediction program P2 also includes a main module m20 that comprehensively controls prediction processing in the prediction device 20, a second performance statistical value acquisition module m21, a second feature data generation module m22, a prediction module m23, and an output module m24. configured. Each module m21 to m24 realizes each function for the second performance statistical value acquisition section 21, the second feature data generation section 22, the prediction section 23, and the output section 24.

Note that the predictive model generation program P1 and the predictive program P2 may be transmitted via a transmission medium such as a communication line, or may be stored in recording media M1 and M2 as shown in FIG. It may be an aspect.

In the predictive model generation device 10, the predictive model generation method, and the predictive model generation program P1 of the present embodiment described above, the performance statistics of one place are added to the performance statistics of another place according to the number of people moving between the places. Feature data is generated by reflecting the The feature data generated in this way constitutes a feature quantity that indicates the characteristics of the activity statistics of one place, and this feature quantity includes the influence of the activity statistics of other places based on the number of people traveling. There is. Then, the predictive model is developed by performing machine learning based on the error between the model output value obtained by inputting the feature data of each location into the predictive model and the correct data consisting of the actual statistical values at the reference time of each location. Constructed. Therefore, it is possible to obtain a prediction model that can accurately predict activity statistics associated with a location.

Further, in a predictive model generation system according to another embodiment, the first feature data generation unit classifies each location into a plurality of groups based on the inter-location relationship information, and assigns a group identifier for identifying the group to each location. Characteristic data for one place may be generated by associating performance statistics of one place with a group identifier associated with one place.

According to the above embodiment, in the feature data of each location, the same group identifier is associated with the performance statistics of each location classified into the same group. Thereby, the feature data of each location having the same group identifier as the one location will be associated with the feature data of the one location via the group identifier. Therefore, in learning the predictive model, the feature data of another location is appropriately reflected in the feature data of one location.

In another aspect of the predictive model generation system, the predictive model includes a neural network, and the predictive model generator generates an error between a model output value obtained by inputting feature data into the neural network and correct data. The parameters of the neural network may be updated based on.

According to the above embodiment, by being assigned the same group identifier, the neural network is trained using the feature data of each place that is associated with each other based on the tendency of the number of people moving, so that the movement of people is It is possible to construct a predictive model of activity statistics that is appropriately considered.

Further, in a predictive model generation system according to another embodiment, the first feature data generation unit generates aggregated statistical values by aggregating performance statistical values of other locations using a predetermined method based on inter-location relationship information. The characteristic data of one place may be generated by combining the performance statistics of one place and the aggregated statistics using a predetermined method.

According to the above configuration, the performance statistics of one place are aggregated with the performance statistics of other places based on the number of people moving between the one place and the other place to generate aggregate statistics. , the characteristic data of one place is generated by combining the aggregated statistics with the performance statistics of one place, so the generated characteristic data is based on the performance of other places according to the number of people moving between the places. Appropriate effects of statistics can be included.

Further, in a predictive model generation system according to another embodiment, the first feature data generation unit uses a graph convolution neural network that uses a graph convolution method that uses each location as a node and that uses inter-location relationships represented by inter-location relationship information as edges between the nodes. The node features of one node corresponding to one location and other nodes corresponding to other locations connected by an edge include a network, and the actual statistical value of each location is the node feature associated with each node. By aggregating the quantities based on the inter-location relationship information corresponding to each edge, an aggregated feature of one node corresponding to the aggregated statistical value for one location is generated, and the node feature of one node and The convolutional feature of one node is generated by multiplying the combined feature generated by combining the aggregated feature of one node with the weight matrix, and the generation of the convolutional feature of each node is performed for one layer. As a convolution process, the convolution feature values generated by a predetermined number of layers of convolution process, in which the convolution feature values of each generated node are used as the node feature values of each node in the next layer convolution process, are combined with the feature data of each location. You can also do it.

According to the above embodiment, feature data for each location is generated by a graph convolutional neural network (GCN) in which each location is a node and the relationship between locations is an edge. Based on the GCN, the performance statistics of other locations adjacent to nodes at one location are aggregated based on the number of people moving between locations corresponding to edges, thereby generating aggregated statistics regarding one location. A convolutional feature amount for a location is generated by combining the aggregated statistic value with the performance statistic value of the location, and further multiplying by a weight matrix. By performing this convolution process over a predetermined number of layers, feature data regarding one location that reflects features of other locations with high accuracy is generated.

In another aspect of the predictive model generation system, the predictive model includes a neural network, and the predictive model generator generates an error between a model output value obtained by inputting feature data into the neural network and correct data. The parameters of the neural network and the weight matrix may be updated based on the above.

According to the above embodiment, the GCN and the prediction model are trained based on the error between the model output value obtained by inputting the feature data obtained as the output of the GCN into the neural network and the correct data. As a result, it is possible to construct a predictive model of activity statistics that appropriately takes into account the movement of people.

In the predictive model generation system according to another embodiment, the first feature data generation unit generates the aggregated statistical value by a weighted average based on the number of people moving in the inter-location relationship information with respect to the actual statistical values of other locations. You can also do it.

According to the above embodiment, it is possible to obtain an aggregated statistical value in which the characteristics of the performance statistical values of other locations are appropriately reflected by using the number of people traveling between locations as weights.

Furthermore, in a predictive model generation system according to another form, the inter-location relationship information includes a first number of people moving from another place to one place and a number of people moving from one place to another place. The first characteristic data generation unit may generate the characteristic data of one place based on the first number of people moving.

According to the above embodiment, in view of the fact that the first number of people moving can have a stronger influence on the activity statistics of one place than the second number of people moving, the first number of people moving is based on the first number of people moving. Characteristic data of the location is generated. This makes it possible to generate characteristic data that more appropriately reflects the number of people moving between locations.

Furthermore, in a predictive model generation system according to another embodiment, the actual statistical value corresponds to the time period to which the reference time belongs in the current period in the previous period which is the unit period before the current period which is the unit period to which the reference time belongs. It is good also as including the performance value of the activity statistics value of a period, and the performance value of the time series activity statistics value of the predetermined period in the past before a reference time.

According to the above configuration, the activity statistics at the base time and the prediction target time reflect the trends in the changes in the activity statistics within the unit period and the trends in the changes in the activity statistics in the most recent predetermined period. In view of this, the actual values of activity statistics in the previous period corresponding to the period to which the reference time belongs in the current period, and the actual values of time-series activity statistics for a predetermined period in the past before the reference time. By including this, it is possible to obtain characteristic data suitable for predicting statistical values.

A prediction system according to one embodiment of the present invention is a prediction system that predicts activity statistics, which are statistics associated with a location, and which predicts performance statistics, which are performance statistics of activity statistics in the past before a prediction target time. By reflecting the performance statistics of other locations other than one of the multiple locations in the performance statistics of one location based on the relationship information between locations, the activity statistics of one location can be calculated. a second feature data generation unit that generates feature data representing characteristics of values, wherein the inter-location relationship information is information indicating the number of people moving between locations; is input into a prediction model to obtain predicted statistical values, and the prediction model uses learning data consisting of feature data of each location and correct data consisting of actual statistical values at a reference time of each location. The prediction unit is generated by machine learning based on the prediction unit.

According to the above embodiment, the characteristic data is generated by reflecting the performance statistics of one place on the performance statistics of another place according to the number of people moving between the places. The feature data generated in this way constitutes a feature quantity that indicates the characteristics of the activity statistics of one place, and this feature quantity includes the influence of the activity statistics of other places based on the number of people traveling. There is. By inputting the feature data of each location into the prediction model, it is possible to obtain predicted statistical values that are predicted activity statistical values at the time of prediction.

Although this embodiment has been described in detail above, it is clear for those skilled in the art that this embodiment is not limited to the embodiment described in this specification. This embodiment can be implemented as modifications and changes without departing from the spirit and scope of the present invention as defined by the claims. Therefore, the description in this specification is for the purpose of illustrative explanation and does not have any restrictive meaning with respect to this embodiment.

Each aspect/embodiment described herein applies to LTE (Long Term Evolution), LTE-A (LTE-Advanced), SUPER 3G, IMT-Advanced, 4G, 5G, FRA (Future Radio Access), W-CDMA. (registered trademark), GSM (registered trademark), CDMA2000, UMB (Ultra Mobile Broadband), IEEE 802.11 (Wi-Fi), IEEE 802.16 (WiMAX), IEEE 802.20, UWB (Ultra-WideBand), The present invention may be applied to systems utilizing Bluetooth (registered trademark), other suitable systems, and/or next-generation systems extended based thereon.

The order of the processing procedures, sequences, flowcharts, etc. of each aspect/embodiment described in this specification may be changed as long as there is no contradiction. For example, the methods described herein present elements of the various steps in an exemplary order and are not limited to the particular order presented.

The input/output information may be stored in a specific location (eg, memory) or may be managed in a management table. Information etc. to be input/output may be overwritten, updated, or additionally written. The output information etc. may be deleted. The input information etc. may be transmitted to other devices.

Judgment may be made using a value expressed by 1 bit (0 or 1), a truth value (Boolean: true or false), or a comparison of numerical values (for example, a predetermined value). (comparison with a value).

Each aspect/embodiment described in this specification may be used alone, may be used in combination, or may be switched and used in accordance with execution. In addition, notification of prescribed information (for example, notification of "X") is not limited to being done explicitly, but may also be done implicitly (for example, not notifying the prescribed information). Good too.

Although the present disclosure has been described in detail above, it is clear to those skilled in the art that the present disclosure is not limited to the embodiments described in the present disclosure. The present disclosure can be implemented as modifications and variations without departing from the spirit and scope of the present disclosure as determined by the claims. Therefore, the description of the present disclosure is for the purpose of illustrative explanation and is not intended to have any limiting meaning on the present disclosure.

Software includes instructions, instruction sets, code, code segments, program code, programs, subprograms, software modules, whether referred to as software, firmware, middleware, microcode, hardware description language, or by any other name. , should be broadly construed to mean an application, software application, software package, routine, subroutine, object, executable, thread of execution, procedure, function, etc.

Additionally, software, instructions, etc. may be sent and received via a transmission medium. For example, if the software uses wired technologies such as coaxial cable, fiber optic cable, twisted pair and digital subscriber line (DSL) and/or wireless technologies such as infrared, radio and microwave to When transmitted from a remote source, these wired and/or wireless technologies are included within the definition of transmission medium.

The information, signals, etc. described in this disclosure may be represented using any of a variety of different technologies. For example, data, instructions, commands, information, signals, bits, symbols, chips, etc., which may be referred to throughout the above description, may refer to voltages, currents, electromagnetic waves, magnetic fields or magnetic particles, light fields or photons, or any of these. It may also be represented by a combination of

Note that terms explained in this disclosure and/or terms necessary for understanding this specification may be replaced with terms having the same or similar meanings.

As used herein, the terms "system" and "network" are used interchangeably.

Further, the information, parameters, etc. described in this specification may be expressed as absolute values, relative values from a predetermined value, or other corresponding information. .

As used in this disclosure, the terms "determining" and "determining" may encompass a wide variety of operations. "Judgment" and "decision" include, for example, judging, calculating, computing, processing, deriving, investigating, looking up, search, and inquiry. (e.g., searching in a table, database, or other data structure), and regarding an ascertaining as a "judgment" or "decision." In addition, "judgment" and "decision" refer to receiving (e.g., receiving information), transmitting (e.g., sending information), input, output, and access. (accessing) (e.g., accessing data in memory) may include considering something as a "judgment" or "decision." In addition, "judgment" and "decision" refer to resolving, selecting, choosing, establishing, comparing, etc. as "judgment" and "decision". may be included. In other words, "judgment" and "decision" may include regarding some action as having been "judged" or "determined." Further, "judgment (decision)" may be read as "assuming", "expecting", "considering", etc.

As used in this disclosure, the phrase "based on" does not mean "based solely on" unless expressly stated otherwise. In other words, the phrase "based on" means both "based only on" and "based at least on."

Any reference to the elements herein, such as "first", "second", etc., does not generally limit the amount or order of those elements. These designations may be used herein as a convenient way of distinguishing between two or more elements. Thus, reference to a first and second element does not imply that only two elements may be employed therein or that the first element must precede the second element in any way.

To the extent that the words "include," "including," and variations thereof are used in this specification or in the claims, these terms are synonymous with the term "comprising." is intended to be comprehensive. Furthermore, the term "or" as used in this specification or in the claims is not intended to be exclusive or.

In this specification, a plurality of devices is also included unless it is clear from the context or technology that only one device exists.

Throughout this disclosure, the plural is intended to be included unless the context clearly dictates otherwise.

DESCRIPTION OF SYMBOLS 1... Prediction model generation system, 2... Prediction system, 10... Prediction model generation device, 11... Movement information acquisition part, 12... Inter-location relationship information generation part, 13... First performance statistics value acquisition part, 14... First Feature data generation unit, 15... Prediction model generation unit, 16... Prediction model output unit, 20... Prediction device, 21... Second performance statistical value acquisition unit, 22... Second feature data generation unit, 23... Prediction unit, 24... Output unit, 31... Movement information storage unit, 32... Inter-location relationship information storage unit, 33... Actual statistical value information storage unit, 34... Prediction model storage unit, M1, M2... Recording medium, m10... Main module, m11... Movement information acquisition module, m12... Inter-location relationship information generation module, m13... First performance statistical value acquisition module, m14... First feature data generation module, m15... Prediction model generation module, m16... Prediction model output module, m20... Main module, m21...Second performance statistical value acquisition module, m22...Second feature data generation module, m23...Prediction module, m24...Output module, P1...Prediction model generation program, P2...Prediction program.

Claims (10)

A predictive model generation system that generates a predictive model for predicting activity statistics that are statistical values associated with a location,
Among the performance statistics that are the performance statistics of the activity statistics in the past before the reference time, the performance statistics of a location other than one of the multiple locations are determined based on the inter-location relationship information. a first feature data generation unit that generates feature data representing the characteristics of the activity statistics of the one location by reflecting the performance statistics of the one location, the first feature data generation unit comprising: a first feature data generation unit, which is information indicating the number of people moving between locations;
A prediction model generation unit that generates the prediction model using learning data consisting of the feature data of each location and correct data consisting of performance statistics at the reference time of each location, a predictive model generation unit that updates parameters of the predictive model based on an error between the model output value obtained by inputting it into the model and the correct data;
A predictive model generation system comprising: The first feature data generation unit includes:
classifying each location into a plurality of groups based on the inter-location relationship information, and associating a group identifier for identifying the group with each location;
generating the characteristic data of the one location by associating a group identifier associated with the one location with performance statistics of the one location;
The predictive model generation system according to claim 1. The predictive model includes a neural network,
The predictive model generation unit updates parameters of the neural network based on an error between the model output value obtained by inputting the feature data into the neural network and the correct data.
The predictive model generation system according to claim 2. The first feature data generation unit includes:
aggregating performance statistics of the other locations using a predetermined method based on the inter-location relationship information to generate aggregate statistics;
generating the characteristic data of the one place by combining the actual statistical value of the one place and the aggregated statistical value using a predetermined method;
The predictive model generation system according to claim 1. The first feature data generation unit includes:
comprising a graph convolutional neural network in which each location is a node and the relationship between locations represented by the inter-location relationship information is an edge between the nodes;
The performance statistical value of each location is used as the node feature associated with each node, and the node feature of another node corresponding to the other location connected by an edge to one node corresponding to the one location are aggregated based on the inter-location relationship information corresponding to each edge, thereby generating an aggregated feature amount of the one node corresponding to the aggregated statistical value regarding the one location, and generating a convolutional feature of the one node by multiplying the combined feature generated by combining the node feature and the aggregated feature of the one node by a weight matrix;
A predetermined number of layers in which the generation of the convolutional feature of each node is one layer of convolution processing, and the generated convolution feature of each node is the node feature of each node in the convolution processing of the next layer. The convolutional feature amount generated by the convolution process is used as feature data of each location,
The predictive model generation system according to claim 4. The predictive model includes a neural network,
The predictive model generation unit updates parameters of the neural network and the weight matrix based on an error between the model output value obtained by inputting the feature data into the neural network and the correct data.
The predictive model generation system according to claim 5. The first feature data generation unit generates the aggregated statistical value by a weighted average using the number of people moving in the inter-location relationship information as a weight with respect to the performance statistical value of the other location.
The predictive model generation system according to any one of claims 4 to 6. The inter-location relationship information includes a first number of people moving from the other place to the one place, and a second number of people moving from the one place to the other place. including,
The first characteristic data generation unit generates characteristic data of the one place based on the first number of people moving.
The predictive model generation system according to any one of claims 1 to 7. The performance statistics are:
Actual values of the activity statistics for a period corresponding to the period to which the reference time belongs in the current period in a previous period that is a unit period before the current period to which the reference time belongs;
actual values of the activity statistics in a time series for a predetermined period in the past before the reference time;
The predictive model generation system according to any one of claims 1 to 8. A prediction system for predicting activity statistics that are statistics associated with a location, the prediction system comprising:
Among the performance statistics that are the performance statistics of the activity statistics in the past before the prediction target time, the performance statistics of a location other than one of the multiple locations are used as inter-location relationship information. a second feature data generation unit that generates feature data representing the characteristics of the activity statistics of the one location by reflecting the performance statistics of the one location based on the relationship between the locations; the information is information indicating the number of people moving between locations, a second characteristic data generation unit;
A prediction unit that inputs the feature data into a prediction model to obtain a predicted statistical value, wherein the prediction model is based on the feature data of each location and correct data consisting of actual statistical values at a reference time of each location. a prediction unit generated by machine learning using learning data;
A prediction system equipped with

Images