JP7499732B2 - Domain information estimation model, apparatus and method including a generator trained with modified event-related information - Google Patents

Description

The present invention relates to a technology for estimating unknown or desired information parts in event and related information, which is information on events distributed within a certain area and events that may be related to them.

In recent years, technology that estimates the movement trajectory of a target person (hereinafter also referred to as a user) in a specified area has attracted much attention. By using this technology, it is possible to provide services such as estimating the user's future movement trajectory, predicting the locations the user will stop at or aim for, and presenting the user with useful information related to those locations in advance.

As an example of such a movement trajectory estimation technology, Patent Document 1 discloses a destination prediction device that can predict the destination of a user even if there is no information on the past movement trajectory. Specifically, this device estimates information regarding the remaining movement of a target user based on the movement trajectory of the target user while he or she is moving, and predicts the destination of the target user based on the information thus estimated and information indicating the number of users moving between areas, which is stored in a movement tendency information storage unit.

Here, the probability of a target user visiting from the departure grid to the destination grid is calculated using the movement trends of other users, as well as the estimated results of the target user's current movement direction, movement distance, and even means of transportation.

For example, Non-Patent Document 1 discloses a technique for estimating a set of trajectories of crowds in a target city for which no trajectory data is available, using a set of trajectories in a source city and road features in each grid. Specifically, this technique performs the above estimation by implementing three steps: mobility intention transfer, OD generation, and path generation.

First, mobility intention transfer involves embedding pairs (OD pairs) of departure grid O and destination grid D in the source city's travel trajectory set into feature space. Here, OD pairs are expressed by the number and categories of POIs (Points of Interest), the number of bus stops and train stations, and the distance of the shortest path between ODs. As a result of this OD pair embedding process, the features of each OD pair, such as "OD pairs for shopping" and "OD pairs for commuting," are learned, and OD pairs with similar purposes are mapped to nearby locations in the feature space.

Next, in OD Generation, OD pair candidates for the Target city are generated. To reduce computational costs, OD pairs are not calculated for all grids, but are calculated only for OD pairs whose longest route is 6 kilometers (km) or less (an empirical threshold). The OD pairs calculated here are also mapped onto the same feature space using the same method as for the Source city above.

Finally, in Path Generation, movement trajectories between candidate OD pairs in the Target city are generated. First, route candidates that can be movement trajectories between candidate OD pairs are extracted in order of shortest route, and one of two OD pairs with a certain amount of overlap is excluded. Next, a DNN (Deep Neural Networks) model is constructed using a set of movement trajectories between OD pairs in Source cities that are adjacent in feature space as training data. Each movement trajectory is represented by the number of road segments and the number of U-turns, and the probability of selecting each movement trajectory is output from the output layer of this DNN model. This makes it possible to determine the probability of occurrence of a movement trajectory set in a Target city where no movement trajectory data exists.

JP 2020-112917 A

Tianfu He, Jie Bao, Ruiyuan Li, Sijie Ruan, Yanhua Li, Li Song, Hui He and Yu Zheng. “What is the Human Mobility in a New City: Transfer Mobility Knowledge Across Cities”, In proceedings of the 2020 International World Wide Web Conference (WWW 2020), pp. 1355-1365, 2020.

However, in the conventional techniques described above, there are still cases where it is difficult to properly estimate the movement trajectory of the target user due to problems with the estimation models that are constructed.

For example, in the technology described in Patent Document 1, when constructing a model for estimating the destination of a target user, a large amount of movement trajectory data of other users in the estimation target area is required as learning data. However, it is generally difficult to obtain such a large amount of movement trajectory data, especially in suburban areas. Therefore, depending on the situation, it may not be possible to estimate the movement trajectory with a high degree of accuracy.

Furthermore, in the technology described in Non-Patent Document 1, one DNN model is constructed for each OD pair, but these DNN models only learn the road features on the trajectory, and only generate short-distance (within 6 km) trajectories as estimated results. Therefore, in many cases, the estimated trajectory is simply the shortest route (within the limits of the road features). In other words, in the technology described in Non-Patent Document 1, there is no room for the selection of a trajectory route that reflects factors other than the road itself, such as the user's tendency to "prefer roads with a lot of people" or "prefer roads with good scenery," in the trajectory estimation process.

The problems mentioned above naturally occur not only when estimating the movement trajectory information of a target user, but also when estimating unknown or confirmed information parts in "event information," which is information related to events distributed within a certain area, or in "related information" that may be correlated or causally related to this information.

The present invention aims to provide an intra-area information estimation model that estimates the information portion of event information and related information within a certain area, which can prepare the training (learning) data required for construction and can also reflect various related information in the estimation, as well as an intra-area information estimation device and method that utilizes said model.

According to the present invention, in intra-domain event/related information, which is a set of event information, which is information related to events distributed within a certain domain, and at least one piece of related information, which is information distributed within the domain and may be correlated or have a causal relationship with the event information, a generator takes as input modified event/related information, which is information obtained by modifying a part of the event information and/or the related information, and generates generated event/related information as an output, the generator being made to function as a trained generator such that the generated event/related information is a restoration of the original intra-domain event/related information,
This generator provides an intra-region information estimation model, which is characterized in that it takes as input the event information and/or related information that serve as premise information for the estimation, and outputs estimation information that is information related to the event information and/or related information that is the subject of estimation.

As an embodiment of the intra-region information estimation model according to the present invention,
The domain information estimation model further functions as a classifier trained to distinguish between the generated event/related information generated by the generator using the change event/related information as input and the domain event/related information that is the source information of the change event/related information;
It is also preferable that the generator is trained so that the generated event and related information it generates cannot be discriminated by the classifier.

In the above embodiment, the generator is also preferably trained with a loss function that shows a monotonically increasing error between the generated event and related information and the original event and related information in the domain, and also shows a monotonically increasing degree of distinctness of the generated event and related information, as determined by the output from the classifier.

Furthermore, it is also preferable that the generator of the intra-domain information estimation model according to the present invention is constructed using an auto-encoder including multiple CNNs (Convolutional Neural Networks) that respectively characterize the event information and the related information.

It is also preferable that the changed event/related information according to the present invention is information in which part of the event information and/or related information is missing.

Furthermore, as a preferred specific example of information handled in the present invention, the event information is movement trajectory information including a group of location positions of certain objects distributed within the region,
It is also preferable that the related information is spatial and semantic information including at least one of information relating to the distribution, arrangement or category of artificial and/or natural objects present in the area, information relating to the distribution, arrangement or category of POIs (Points of Interest) set in the area, information representing the shortest route between specified location ranges in the area, information relating to the number of the objects in each location range in the area, and information relating to the number of posts of a networking service in each location range in the area or the analysis results of the posts.

In the above specific example, it is also preferable that the generator takes the spatial and semantic information as the premise information as input and outputs the estimated information related to the movement trajectory information that is the estimation target, or that the generator takes the movement trajectory information and the spatial and semantic information that is not the estimation target as input as the premise information and outputs the estimated information related to at least one of the spatial and semantic information that is the estimation target.

The present invention also provides an intra-area information estimation device that uses the intra-area information estimation model described above to estimate information related to the event information and/or related information that is the estimation target.

According to the present invention, further comprising:
a step of training a generator that uses modified event/related information, which is information obtained by modifying a part of event information and/or related information, as input to generate generated event/related information in intra-domain event/related information, which is a set of event information, which is information related to events distributed within a certain domain, and at least one related information, which is information distributed within the domain and may be correlated or causally related to the event information, so that the generated event/related information is a restoration of the original intra-domain event/related information;
The present invention provides a method for estimating information within a region on a computer, the method comprising the steps of inputting the event information and/or related information that serve as premise information for estimation to the trained generator, and outputting estimated information that is information related to the event information and/or related information that are the subject of estimation.

The present invention provides an intra-area information estimation model that estimates the information portion of event information and related information within a certain area, and that can prepare the training (learning) data required for construction and can also reflect various related information in the estimation, as well as an intra-area information estimation device and method that utilizes the model.

1 is a schematic diagram showing an embodiment of a region information estimation model according to the present invention and an embodiment of a region information estimation device according to the present invention; FIG. 2 is a schematic diagram for explaining an embodiment of intra-area event/related information according to the present invention. FIG. 13 is a schematic diagram for explaining an embodiment of an intra-region information estimation process in a generator according to the present invention.

The following describes in detail an embodiment of the present invention with reference to the drawings.

[Region information estimation model]
FIG. 1 is a schematic diagram showing an embodiment of an intra-region information estimation model according to the present invention, and an embodiment of an intra-region information estimation device according to the present invention.

The region information estimation model 1 of this embodiment shown in FIG.
(a) "Event information" (in FIG. 1, movement trajectory information including a group of locations of the person) which is information related to an event (in FIG. 1, the movement of the person) distributed within a certain area (in FIG. 1, a certain region),
(b) This model is capable of estimating an information portion of the "event/related information within the region" which is a pair of at least one "related information" that is information that is distributed within the region and may be correlated or causally related to the "event information" (movement trajectory information), such as an unknown or desired movement trajectory portion.

Here, in the embodiment in which the “event information” in the above (b) is movement trajectory information,
(b1) Spatial information, such as information on the distribution or layout of roads, houses, and rivers existing in the area (region),
(b2) This is semantic information, such as information regarding the number of objects (e.g., people) in each location range within the area. In any case, the intra-area information estimation model of the present invention makes it possible to select various pieces of information that may be related and set them as "related information."

Also in FIG. 1, the area information estimation model 1 is
(c) "Changed Event/Related Information" is information obtained by changing (e.g., deleting) a part of "Event Information", a part of "Related Information", or a part of "Event Information" and "Related Information" in "Event/Related Information in the Domain".
The system is provided with a generator 12 that receives the above as input and generates "created event and related information" as output (the computer functions as the generator 12). Note that the "changed event and related information" in (c) above is generated by a defect generator (change generator) 11 in this embodiment.

The generator 12 is trained so that the "generated events and related information" it generates is a restoration of the original "domain events and related information," in other words, the "domain events and related information" before the changes were made to the input "changed events and related information." In other words, the generator 12 corresponds to a generator in so-called GANs (generative adversarial networks).

Next, the thus trained generator 12 performs the intra-region information estimation process. Specifically, the following is performed on the generator 12:
(d1) “Event information”, “related information”, or “event information” and “related information” (hereinafter abbreviated as “event information” and/or “related information”) that are premise information for the estimation (known or confirmed)
With input,
(d2) Outputting estimated information that is information related to the (unknown or desired to be confirmed) "event information" and/or "related information" that is the subject of estimation. For example, information on the distribution of actual roads, houses, and rivers in a certain area (related information) and information on the distribution of actual people (related information) may be input, and movement trajectory information (event information) of a certain person in the said area may be output.

It is also possible to input information on the actual movement trajectory of a person in a certain area (event information) and information on the distribution of actual roads, houses, and rivers (related information) and output information on the distribution of people in that area (related information). It is also possible to input information on the actual distribution of roads, houses, and rivers in a certain area (related information) and other related information and output information on the movement trajectory of a person in that area (event information) and information on the distribution of people in that area (related information). It is also possible to include information with some parts missing from the event information or related information in the input and output information with the missing parts generated (complemented).

In this way, according to the intra-area information estimation model 1, when estimating the information portion of the "internal area events and related information", the information portion can be estimated with high accuracy by applying the original information restoration function of at least the Generator in the GAN to the intentionally generated "changed events and related information". It is also possible to reflect various related information in this estimation. For example, as will be described in detail later, it is possible to select "related information" that can reflect a person's tendency to "prefer roads with a lot of people" or "prefer roads with good scenery" in the estimation results of the person's movement trajectory information.

In addition, as related information, it is possible (and more preferably easy) to select information that allows for the preparation of training (learning) data necessary for model construction, so that the intra-region information estimation model 1 makes it possible to reliably prepare the training data necessary for model construction.

For example, when estimating the movement trajectory information of a person in a certain area, it is possible to use the movement trajectory information of other groups of people (crowds) in the same area as "related information." However, in cases such as suburban areas, it is often difficult to collect the large amount of movement trajectory information required to build a model. Even in such cases, the area information estimation model 1 makes it possible to select and set "related information" that is easy to obtain, other than the movement trajectory information of other groups of people (crowds).

Of course, the "areas" and "event information" according to the present invention are not limited to the (two-dimensional) regions and movement trajectory information described above. In particular, the "event information" can be set to information related to social and economic events that are distributed in a certain area and that are in high demand for prediction, such as the crime rate in each district or the distribution of electric vehicle charging points.

Furthermore, it is also possible to adopt, for example, a real one-dimensional or three-dimensional space, or a virtual (theoretical) one-dimensional or two-or-more-dimensional space as the "area," and adopt information related to events that show distribution in the adopted space as the "event information." For example, a two- or three-dimensional culture area in a bacterial incubator can be used as the "area," and bacterial cell concentration information within the culture area related to the bacterial growth "event" can be used as the "event information," and further information on the distribution of nutrients and antibiotics within the culture area can be used as the "related information." Also, in a virtual one-dimensional "area" that has only a time axis, it is also possible to use information on time changes in stock prices related to the "event" of buying and selling stocks as the "event information," and information related to various incidents and phenomena that may affect stock prices as the "related information."

Furthermore, in a virtual two-dimensional "area" having a time axis and an axis for position on a road, information on changes over time in the distribution of automobiles on that road related to automobile driving events can be regarded as "event information," and information on the occurrence and disappearance of events that may affect automobile driving, such as accidents, construction work, and even the status of traffic signals on that road, can be regarded as "related information."

[Model configuration]
An embodiment of the region information estimation model according to the present invention will be described in more detail below. Note that in the embodiment shown below, the region, the event information, and the related information are respectively a predetermined region, movement trajectory information of a target person to be estimated in the predetermined region, and spatial information and semantic information in the predetermined region, but as described above, the present invention is not limited to these.

Also according to FIG. 1, the region information estimation model 1 of the present embodiment is as follows:
The system includes (A) a defect generation unit (change generation unit) 11, (B) a generator 12, and (C) a discriminator 13. Of these, the generator 12 and the discriminator 13 correspond to a generator and a discriminator of a GAN, respectively, and in this embodiment, as described in detail later, both are constructed using neural network algorithms including CNNs (Convolutional Neural Networks).

As described above, the intra-area information estimation model 1 performs intra-area information estimation processing by processing "intra-area events and related information," which is a combination of movement trajectory information and spatial and semantic information (related information). Here, we will explain this "intra-area events and related information."

Figure 2 is a schematic diagram for explaining one embodiment of the area event and related information according to the present invention.

First, FIG. 2(A) shows an example of how movement trajectory information is expressed as event information. Specifically, in the movement trajectory information in FIG. 2(A), a specific area is divided into a group of grids (e.g., a rectangle of 250m x 250m), and in each grid, the movement trajectory (changes in location) of a target person is expressed by setting the value of the grid to 1 if the target person's location during a specific period of time is within that grid, and 0 if not. In this way, if the grids are considered to be pixels, the movement trajectory information can be considered as binary image information in which the pixel values are 0 or 1.

In the example of FIG. 2(A), in a group of adjacent grids with a value of 1, the grid at one end is a departure grid including the target person's departure point, and the grid at the other end is a destination grid including the target person's destination point. However, the movement trajectory information can also be information in which one or both of the departure grid and destination grid are not expressed (set).

The location of the target person can be, for example, the location of the communication terminal carried by the target person, and a telecommunications carrier can obtain information related to this location. Specifically, the location of this communication terminal can be (a representative point of) the area covered by the connected base station, and can also be determined from the results of positioning using a base station positioning method or a GPS (Global Positioning System) device built into the communication terminal.

Next, Figure 2 (B) shows an example of the expression of mesh number of people information as semantic information (related information). Specifically, in the mesh number of people information of Figure 2 (B), the same predetermined area as the above-mentioned movement trajectory information is divided into the same grid group, and in each grid, the mesh number of people (when 1 grid is 1 mesh) is expressed by taking the average number of people (held by communication devices) located within that grid during the predetermined period as the value of that grid. In this way, if the grids are considered to be pixels, the mesh number of people information can also be considered as image information that takes on multiple pixel values.

The grids resulting from dividing the specified area may be of different sizes and shapes than those described above. It is also possible to set the grids for the movement trajectory information (event information) and the grids for the mesh number of people information (related information) to different sizes and shapes, although this may affect the accuracy of subsequent estimation.

As explained above, when the event information is movement trajectory information, the related information is not limited to the mesh number of people information described above. In fact, it is possible to use a wide variety of information as related information, so long as it is information that can be correlated or causally related to the movement trajectory information and is available in sufficient quantities as training data.

For example, such related information includes:
<Spatial information>
(a) information on the distribution, arrangement, or category of artificial and/or natural objects, such as roads, houses, and rivers, existing within the specified area, within a group of grids (grid coverage information);
(b) information on the distribution, arrangement, or category of POIs (Points of Interest) such as stations, leisure and sports facilities, and stores set in the specified area within the grid group, and (c) information representing the shortest route between a specified location range (e.g., between two set grids) within the specified area (e.g., information consisting of each grid value taking the value of 0 or 1),
<Semantic information>
The related information can be at least one of the following: (a) information on the number of predetermined objects such as people and cars in each location range (each grid) in the specified area (e.g., the mesh number information described above), and (e) information on the number of posts on networking services such as SNS (Social Networking Service) and mini-blogs in each location range (each grid) in the specified area, or analysis results of the posts (e.g., frequently occurring words). Incidentally, any of the above information (a) to (e) can be regarded as image information in which grids are regarded as pixels, as shown in Figures 2(A) and 2(B).

The structure of the intra-domain events and related information has been explained above. For example, if D pieces of different related information are provided for this intra-domain event and related information, and the event information and each of the D pieces of related information are partitioned into H×W grids, it is possible to regard this intra-domain event and related information as an H×W×(D+1)-dimensional (rank 3) tensor quantity.

Returning to FIG. 1, the defect generation unit 11 generates changed event/related information, which is information in which part of the movement trajectory information (event information), part of the spatial information (related information), and/or part of the semantic information (related information) is missing from the intra-area event/related information described above.

Specifically, the defect generation unit 11 can replace one or more grid values (numerical values written in grids such as those shown in FIG. 2(A) and FIG. 2(B)) selected randomly (or according to a predetermined rule) in the movement trajectory information, spatial information, or semantic information with a "value that specifies a defect." For example, in the movement trajectory information, one of the grids with a grid value of 1 (indicating that a movement trajectory is included) may be randomly selected, and the grid value of the selected grid may be replaced with 2 (specifying a defect).

As a modification, the defect generation unit 11 may function as a modification generation unit and replace one or more grid values selected randomly (or under a predetermined rule) in the movement trajectory information, spatial information, or semantic information with another possible grid value. For example, in the movement trajectory information, one of the grids with a grid value of 1 may be randomly selected, and the grid value of the selected grid may be replaced with 0 (indicating that the grid does not include a movement trajectory). Furthermore, it is also possible to increase the number of POIs in a certain grid by one and add one category (a value specifying) for this POI.

Also in FIG. 1, the generator 12 is a functional component that takes the change event/related information generated by the defect generation unit (change generation unit) 11 as input and generates generated event/related information as output, where the generated event/related information is trained to restore the domain event/related information that is the original information of the input change event/related information.

Specifically, the generator 12 corresponds to the Generator of the GAN as described above, and can be configured using various machine learning algorithms that can configure this Generator. Here, the generator 12 of this embodiment assumes that the input change event/associated information includes D pieces of mutually different associated information (spatial information/semantic information), and generates
(12a) (D+1) CNNs that take the movement trajectory information of change events and related information and D related information as input, extract and output the features, in other words, feature the movement trajectory information and the D related information, respectively;
(12b) a combination layer that combines the output data from the (D+1) CNNs;
(12c) An autoencoder that uses the output data from the combined layer of (12b) above as input and includes (D+1) inverse CNNs that output the movement trajectory information of the generated event and related information, and D related information.

In other words, in this embodiment, the generation unit 12 regards the input trajectory information and the D related information as image information (image data), extracts their image features, and then performs a process of restoring the original (before missing or altered) trajectory information and the D related information (a process of repairing missing information if missing information has been lost) all at once. This means that the calculation cost for the estimation process is significantly reduced, for example, compared to the trajectory estimation technology described in Non-Patent Document 1, which constructs one DNN model for each OD pair.

Here, the training of the generator 12 as an autoencoder will be described. Basically, the generator 12, as described above,
(a) generated events and related information to be output;
(b) Training is performed so that the difference (error) between the input changed event/related information and the original information, that is, the domain event/related information, is as small as possible. Specifically, the training process can be performed using a loss function MSE_Loss, whose function value is the mean square error (MSE) between the grid values (pixel values) of the grid of the generated event/related information in (a) above and the grid values (pixel values) of the corresponding grid in the domain event/related information in (b) above.

When training the generator 12 using only this loss function MSE_Loss, the discriminator 13 described below is not necessary, and therefore the intra-region information estimation model 1 can perform intra-region information estimation processing using only the generator 12. However, in this embodiment, in order to ensure the inherent estimation accuracy of the GAN, the output from the discriminator 13 is also used to train the generator 12.

Also in FIG. 1, the discriminator 13
(a) generated event and related information generated by a generator 12 using certain changed event and related information as input;
(b) A functional component trained to input the domain event and related information, which is the source information of the change event and related information in (a) above, and to discriminate between these two pieces of information.

Specifically, the classifier 13 corresponds to a GAN discriminator as described above, and can be configured using various machine learning algorithms that can configure this discriminator. Here, the classifier 13 of this embodiment assumes that the input intra-domain event/related information and change event/related information each contain D pieces of mutually different related information (spatial information/semantic information), and
(13a) (D+1) CNNs that each take the movement trajectory information of events and related information in the domain and D related information as input, extract and output the features;
(13b) (D+1) CNNs that respectively input the movement trajectory information of the generated event and related information and D related information, extract and output the feature amounts;
(13c) a combination layer that combines the output data from the CNNs of (13a) and (13b) above;
(13d) An output layer which outputs, based on the output data from the combination layer in (13c) above, information relating to distinguishing between the generated event/related information and the domain event/related information, for example, the probability that the two are the same information (i.e., the generated event/related information is information that has been restored to its original state).

In addition, the training process of the classifier 13 uses, for example, a known binary cross entropy loss as a loss function for training a binary classification model, as follows:
This can be implemented by learning a large number of first training data which are pairs of intra-domain event/related information (input to the CNN in (13a) above) and intra-domain event/related information identical to this information (input to the CNN in (13b) above) labeled as “same information”, and a large number of second training data which are pairs of intra-domain event/related information (input to the CNN in (13a) above) and generated event/related information generated from this information by generator 12 (input to the CNN in (13b) above) labeled as “different information”.

Here, the output from the output layer (13d) of the classifier 13 is the probability _pC that the generated event/related information and the domain event/related information input in this embodiment are the same information. In this embodiment, the above-mentioned generator 12 is trained using this output (probability _pC ).

Specifically, the training process of the generator 12 is, for example, the following equation (1): Loss = α MSE_Loss + (1 - α) GAN_Loss
GAN_Loss＝1－p _C
This can be implemented using a loss function Loss expressed by: where α is a weighting parameter set to a predetermined value. In addition, GAN_Loss (=1-p _C ) in the above formula (1) is interpreted as a quantity indicating the degree to which the generated events and related information are different, determined by the output (probability p _C ) from the discriminator 13. Therefore, by using a loss function including such GAN_Loss, the generator 12 is trained so that the generated events and related information generated by the generator 12 cannot be distinguished by the discriminator 13, that is, so that the original event and related information in the domain can be regarded as restored from the perspective of the discriminator 13.

It should be noted that the loss function Loss is not limited to the form of the above formula (1). In any case, the loss function Loss for training the generator 12 is as follows:
(a) The error (e.g., the sum of the errors for each grid) between the generated event and related information generated by the generator 12 and the original event and related information within the domain is monotonically increasing, and
(b) It is preferable to adopt a monotonically increasing degree of distinctness of generated events and related information, which is determined by the output from the classifier 13 .

As a modification, the training process of the generator 12 can be performed using only GAN_Loss, or using only the loss function (b) above that "indicates a monotonically increasing degree of distinctness of the generated events and related information." In this case, the training process does not require the computational cost of the loss function MSE_Loss.

Similarly, in FIG. 1, the generator 12 that has been subjected to the training process described in detail above performs the following as an intra-region information estimation process:
(a) "Prerequisite information for estimation" which is event information and/or related information as premise information
With input,
(b) "Estimated information" is information related to "event information and/or related information" as the subject of estimation
Output.

Specifically, in this embodiment, the generator 12:
(A) Inputting related information (spatial information, semantic information) as “prerequisite information for estimation” and outputting “estimated information” related to the trajectory information to be estimated, or
(i) As "prerequisite information for estimation," movement trajectory information and related information (spatial information, semantic information) that is not the subject of estimation are input, and "estimated information" relating to at least one of the related information (spatial information, semantic information) that is the subject of estimation is output.

Figure 3 is a schematic diagram illustrating one embodiment of the region information estimation process in the generator 12.

In the embodiment shown in FIG. 3, the trained generator 12 (trained in a training unit 93 described below)
(a) trajectory information in which a starting grid and a destination grid including a starting point and a destination point, respectively, are specified and the remaining grids are missing;
(b) Prerequisite information for estimation, which is a pair of actual related information (spatial information, semantic information), is input,
(c) Outputting estimated information including movement trajectory information representing an estimated movement trajectory when the departure grid and destination grid in (a) above are set as the departure range and destination range, respectively (specifically, each grid value of the grid group including the estimated location group is set to 1), or which is the movement trajectory information itself.

In this way, the trained generator 12 can estimate the movement trajectory of, for example, a person who is the same as a person in the training data in a new area by specifying a start grid and a destination grid in the new area that is different from the area related to the training data, and related information representing the actual positions and distribution of roads, houses, and rivers in the new area, and the actual number of people in each mesh, etc.

For example, by using a generator 12 trained with training data related to the area where a person lived before moving, and specifying a starting grid including the location of the person's new home in the new area to which the person has moved, and a destination grid including the location of the person's new workplace that will be the person's commute destination, and further specifying related information for the new area to which the person has moved, it is possible to generate and estimate a new commuting route for the person in the new area.

Here, by using relevant information such as the time average number of people in each mesh (mesh number of people) or the number of buildings along a road that are above a certain height, it becomes possible to estimate a more plausible movement trajectory (commuting route) for the person that reflects, for example, the person's tendency to "prefer roads with a lot of people" or "prefer roads with good scenery."

In the example described above, the generator 12 is an estimator dedicated to a specific person related to the training data, and as described above, it is possible to reflect the preferences and tendencies (in route selection) of that person in the estimation result. However, as a modification, the generator 12 may be trained with training data related to a group of people who belong to a specific organization/group or have specific attributes, and the trained generator 12 may be used to estimate the movement trajectory of the person. It is also possible to estimate the movement trajectory of any person using a generator 12 trained with training data related to an unspecified group of people.

Furthermore, for example, it is possible to adopt, as the movement trajectory information of a certain area, movement trajectory information representing the location group of a large number of people (communication terminals owned by the people), i.e., the movement trajectory group of a group of people (crowd), and specify as including in the estimation premise information (related information) the road distribution (different from the road distribution of the area) in a new area different from the area related to the training data, and estimate the movement trajectory information (road usage information) of the crowd in this new area. In this case, it becomes possible to estimate information related to the trajectory in accordance with the actual movement of the crowd, without being limited to the movement trajectory estimation of a short distance (within 6 km) as in the movement trajectory estimation technology described in Non-Patent Document 1.

In the above example, a movement trajectory in a different area from the area related to the training data is estimated, but it is of course possible to estimate a movement trajectory in the same area. For example, by specifying related information that is the arrangement, distribution, and category information of POIs including newly established POIs (e.g., new commercial facilities) that were not included in the training data, a movement trajectory that reflects the establishment of new POIs in the same area, such as a commuting route, may be estimated. In this way, the generator 12 makes it possible to estimate a new movement trajectory when the environment of the person to be estimated changes, such as moving, traveling, opening a new store, or even the occurrence of construction or a disaster.

Furthermore, as another embodiment of the intra-area information estimation process, movement trajectory information can be included in the estimation premise information to estimate at least one of the related information. For example, the movement trajectory information representing the group of movement trajectories of a crowd can be fixed, while the generator 12 can be constructed using as training data change event/related information in which part of the related information, such as category information of POIs present in each grid and information on the number of people in each grid (mesh number of people), is missing (or changed). In this case, the generator 12 can estimate the number of people in each grid (mesh number of people) and the category of POI in each grid from a large amount of movement trajectory information in a certain area.

Furthermore, it is possible to leave out a portion of the movement trajectory information included in the premise information for estimation (as unknown information or information to be confirmed), and restore and estimate this missing portion of the movement trajectory information in parallel with estimating the mesh number of people and POI categories.

In any case, the domain information estimation model 1 of these embodiments applies the original information restoration function of the GAN to the intentionally generated changed event and related information, making it possible to restore and estimate unknown or confirmed information parts in the event and related information (prerequisite information for estimation). In addition, as the event information and related information in the event and related information (prerequisite information for estimation), a wide variety of information can be selected and used to prepare the training data required for construction. As a result, by appropriately selecting and adjusting the event information and related information, it is possible to estimate the desired domain information with higher accuracy.

[Area information estimation device, program, and method]
Returning to Figure 1, we will now explain the regional information estimation device 9, which is equipped with the regional information estimation model 1 as described in detail above and is capable of outputting estimation information relating to the event information and/or related information that is the subject of estimation (e.g., travel route information, category information of POIs present in each grid, etc.) as estimation results using this regional information estimation model 1.

As shown in FIG. 1, the intra-area information estimation device 9 includes an input unit 91, an event/related information generation unit 92, a training unit 93, an intra-area information estimation unit 94, and an output unit 95. Of these, the event/related information generation unit 92, the training unit 93, and the intra-area information estimation unit 94 are the main units that implement one embodiment of an intra-area information estimation method according to the present invention, and can also be considered as functions of a processor/memory that stores one embodiment of an intra-area information estimation program according to the present invention.

For this reason, the area information estimation device 9 may be a dedicated device for area information estimation, but it may also be, for example, a cloud server, a non-cloud server device, a personal computer (PC), a notebook or tablet computer, or a smartphone, etc., equipped with the area information estimation program according to the present invention.

Also in FIG. 1, the input unit 91 of the area information estimation device 9 has a communication function and receives and acquires information for generating area event and related information. For example, using a specified API (Application Programming Interface), it may acquire time series information on the location of a communication terminal (or a person carrying the communication terminal) from, for example, a communication management server of an externally installed communication carrier, acquire map information including information on POIs such as roads, houses, rivers, and even buildings and various facilities in the target area from, for example, a map information management server of a specified company, or acquire time series information on the number of posts in each area and information on the analysis results of posts (for example, frequently occurring words) from, for example, a post management server of a networking service provider.

The event and related information generation unit 92 determines information for each grid in a group of grids that divides the target area, which has been set in advance, from the various information about the target area acquired by the input unit 91, as described above, and thereby generates intra-area event and related information for the target area.

The training unit 93 generates a large number of changed events and related information from the large number of generated intra-domain events and related information, and constructs an intra-domain information estimation model 1 using pairs of the generated changed events and related information and the intra-domain events and related information that is the original (missing or pre-changed) information as training (learning) data. In addition, the constructed intra-domain information estimation model 1 may be updated, for example, by the latest training data.

The area information estimation unit 94 generates prerequisite information for estimation (e.g., specifying the starting grid and destination grid as shown in FIG. 3) according to the area information to be estimated as instructed by the device user (via the input unit 91), inputs this prerequisite information for estimation into the constructed and updated area information estimation model 1, and obtains estimation information (e.g., complementing and estimating the movement trajectory connecting the starting grid and destination grid as shown in FIG. 3) as output from this area information estimation model 1.

The output unit 95 transmits the estimated information received from the area information estimation unit 94 to an external information processing device if it has a communication function, or displays it on a display if it has a display. For example, the estimated movement trajectory may be highlighted on a map image of the area, the category of the estimated POI may be presented in the form of a speech bubble from the position of the POI, and the estimated number of people in a mesh may be expressed by shading on the map image.

As described above in detail, according to the present invention, when estimating information portions of events and related information within a domain, the original information restoration function of at least a Generator in the GAN is applied to the generated changed events and related information, making it possible to estimate the information portions with high accuracy. In addition, various information that makes it possible to prepare training (learning) data necessary for model construction can also be reflected in this estimation.

Furthermore, for example, by adopting the movement trajectory information of groups of people or vehicles as the event information of the present invention, it is possible to predict the flow of people and traffic of vehicles when an event occurs in an urban area, when a disaster occurs, or when a state of emergency is declared, and to implement appropriate control of the flow of people and traffic based on the prediction results. In this way, the present invention is also widely used in situations where the steady-state fluctuation factors of various events in smart cities, which are currently a hot topic, are to be estimated.

The present invention also makes it possible to determine the layout and distribution of social infrastructure (including transportation) and commercial facilities that are easy for residents to access and that do not cause congestion for residents who access them, and to determine the appropriate layout and distribution of power supply sources such as cogeneration and solar panels and various power consumers to realize an appropriate power intake and supply system. In other words, the present invention makes it possible to contribute to Goal 11 of the United Nations' Sustainable Development Goals (SDGs), "Make cities inclusive, safe, resilient and sustainable."

With respect to the various embodiments of the present invention described above, various changes, modifications, and omissions within the scope of the technical ideas and viewpoints of the present invention can be easily made by a person skilled in the art. The above description is merely an example and is not intended to be restrictive in any way. The present invention is limited only by the scope of the claims and their equivalents.

1 Region information estimation model 11 Missing part generation unit (change generation unit)
12 Generator 13 Classifier 9 Region information estimation device 91 Input unit 92 Event/related information generation unit 93 Training unit 94 Region information estimation unit 95 Output unit

Claims (9)

a generator that receives input of event information, which is information related to events distributed within a certain domain, and at least one piece of related information, which is information distributed within the domain and may be correlated or have a causal relationship with the event information, and generates generated event/related information as output, the generator being trained so that the generated event/related information is a restoration of the original event/related information within the domain;
The generator receives the event information and/or related information that serve as premise information for the estimation as input, and outputs estimated information that is information related to the event information and/or related information that are the subject of estimation. The computer is further made to function as a classifier trained to distinguish between the generated event/related information generated by the generator using the change event/related information as input and the event/related information in the domain that is the source information of the change event/related information;
2. The domain information estimation model according to claim 1, wherein the generator is trained so that the generated event and related information generated by the generator itself cannot be discriminated by the classifier. The intra-domain estimation model according to claim 2, characterized in that the generator is trained with a loss function that shows a monotonically increasing error between the generated event and related information and the original intra-domain event and related information, and also shows a monotonically increasing degree of distinctness of the generated event and related information, as determined by the output from the classifier. The intra-region information estimation model according to any one of claims 1 to 3, characterized in that the generator is constructed using an auto-encoder including multiple convolutional neural networks (CNNs) that respectively quantify the event information and the related information. The area information estimation model according to any one of claims 1 to 4, characterized in that the changed event/related information is information in which part of the event information and/or the related information is missing. The intra-area information estimation model according to any one of claims 1 to 5, characterized in that the event information is movement trajectory information including a group of locations of certain objects distributed within the area, and the related information is spatial and semantic information including at least one of information related to the distribution, arrangement, or category of artificial objects and/or natural objects existing within the area, information related to the distribution, arrangement, or category of POIs (Points of Interest) set within the area, information expressing the shortest route between predetermined location ranges within the area, information related to the number of the objects in each location range within the area, and information related to the number of posts of a networking service in each location range within the area or the analysis results of the posts. The generator includes:
Using the spatial and semantic information as the premise information as an input, outputting the estimated information related to the trajectory information to be estimated; or
The intra-area information estimation model according to claim 6, characterized in that the movement trajectory information and the spatial/semantic information that is not the target of estimation are input as the premise information, and the estimation information related to at least one of the spatial/semantic information that is the target of estimation is output. An area information estimation device characterized by estimating information related to the event information and/or related information to be estimated using an area information estimation model described in any one of claims 1 to 7. a step of training a generator that uses modified event/related information, which is information obtained by modifying a part of event information and/or related information, as input to generate generated event/related information in intra-domain event/related information, which is a set of event information, which is information related to events distributed within a certain domain, and at least one related information, which is information distributed within the domain and may be correlated or causally related to the event information, so that the generated event/related information is a restoration of the original intra-domain event/related information;
and inputting the event information and/or related information that are premise information for estimation to the trained generator, and outputting estimated information that is information related to the event information and/or related information that are the subject of estimation.

Images