Program, apparatus, and method for evaluating simulated movement trajectories

Abstract

From the perspective of anonymity for measured movement trajectories, we provide a program for evaluating simulated movement trajectories. [Solution] In the training phase, the measured movement trajectory of time-series position coordinates and point identifiers is used as training data. A position coordinate input layer converts the time-series position coordinates of the measured movement trajectory into a position coordinate embedding representation, a point identifier input layer converts the time-series point identifiers of the measured movement trajectory into a point identifier embedding representation, a hidden layer fully connected to both the position coordinate input layer and the point identifier input layer, and a position coordinate output layer fully connected to the hidden layer and estimating the time-series position coordinates. This enables the system to function. The position coordinate output layer is then trained by backpropagation to output the time-series position coordinates in the measured movement trajectory, along with the hidden layer, the point identifier input layer, and the position coordinate input layer.

Description

Technical Field

[0001] The present invention relates to a technology for evaluating the quality of pseudo-generated movement trajectories. Background Art

[0002] In recent years, navigation applications and traffic information providing services for users, and operation management systems for vehicle dispatching and delivery have become commonplace. Among these various services, machine learning models trained using movement trajectories of people, vehicles and the like are used to achieve highly accurate operation.

[0003] Conventionally, there is a technology that uses a diffusion model to generate pseudo movement trajectories from actually measured movement trajectories (see, for example, Non-Patent Document 1). According to this technology, a large amount of pseudo movement trajectory data can be generated from a small amount of actually measured movement trajectory data using a diffusion model. By training a machine learning model using these pseudo movement trajectories as teacher data, the model can also be used for various services. According to the technology of Non-Patent Document 1, pseudo movement trajectories corresponding to actually measured movement trajectories are evaluated and generated in the following aspects. (1) Difference in density of geographical distribution in movement trajectories (2) Difference in positions of start point and end point in movement trajectories (3) Difference in distance in movement trajectories (4) Difference in similarity of trajectory patterns in movement trajectories Prior Art Documents Non-Patent Documents

[0004] Non-Patent Document 1 Yuanshao Zhu et al., “DiffTraj: Generating GPS Trajectory with Diffusion Probabilistic Model,” arXiv: 2304.11582, Oct 2023., [online], [Retrieved February 3, 2020], Internet<URL:https: / / arxiv.org / abs / 2304.11582> [Overview of the project] [Problems that the invention aims to solve]

[0005] However, according to the technology described in Non-Patent Document 1, a pseudo-movement trajectory is generated from the standpoint of "validity" of the measured movement trajectory. If the pseudo-movement trajectory is copied from the measured movement trajectory, it will be evaluated as having "high validity" as a movement trajectory. On the other hand, when applying movement trajectories to third-party services, the aspect of "anonymization" of the measured movement trajectory is essential for privacy protection. In other words, the simulated movement trajectory must not only be evaluated as having "high validity" compared to the measured movement trajectory, but also as having "high anonymity."

[0006] Therefore, the present invention aims to provide a program, an evaluation device, and a method that can evaluate a simulated movement trajectory from the viewpoint of anonymity of the measured movement trajectory. [Means for solving the problem]

[0007] According to the present invention, a program for operating a computer mounted on a device for evaluating movement trajectories, For the training phase, the measured movement trajectory of time-series position coordinates and point identifiers will be used as training data. A position coordinate input layer that converts the time-series position coordinates of the measured movement trajectory into a position coordinate embedding representation, A location identifier input layer that converts location identifiers from the time series of measured movement trajectories into a location identifier embedded representation, A hidden layer that is fully connected to both the position coordinate input layer and the point identifier input layer, A position coordinate output layer is connected to the hidden layer via a fully connected array to estimate the time-series position coordinates. and make it work The position coordinate output layer is trained by backpropagation to output time-series position coordinates in the measured movement trajectory, consisting of the position coordinate output layer, a hidden layer, a point identifier input layer, and a position coordinate input layer. It is characterized by making the computer function in such a way.

[0008] According to other embodiments of the program of the present invention, Regarding the training phase, A location identifier output layer is connected to the position coordinate output layer via a fully connected interface, and estimates the time-series location identifier. And make it work even better. The location identifier output layer is trained by backpropagation to output time-series location identifiers in the measured movement trajectory, along with the location coordinate output layer, the hidden layer, the location identifier input layer, and the location coordinate input layer. It is also preferable to make the computer function in this way.

[0009] According to other embodiments of the program of the present invention, Regarding the training phase, A travel time output layer is connected to the location identifier output layer in a fully connected manner to estimate travel time. And make it work even better. The movement time output layer is trained by backpropagation to output the movement time in the measured movement trajectory. This training includes the movement time output layer, the point identifier output layer, the position coordinate output layer, the hidden layer, the point identifier input layer, and the position coordinate input layer. It is also preferable to make the computer function in this way.

[0010] According to other embodiments of the program of the present invention, For the evaluation phase, the target data will be the pseudo-movement trajectory of time-series position coordinates and point identifiers. The position coordinate input layer receives the time-series position coordinates of the pseudo-movement trajectory, The point identifier input layer receives time-series point identifiers of a pseudo movement trajectory, the position coordinate output layer outputs time-series position coordinates estimated for the pseudo movement trajectory, position distance calculation means for calculating a distance between the time-series position coordinates output from the position coordinate output layer and the time-series position coordinates of the pseudo movement trajectory; and evaluation means for calculating an evaluation value based on said distance from said position distance calculation means it is also preferable to cause a computer to function as said means.

[0011] According to another embodiment of the program of the present invention, the evaluation means from the perspective of anonymity, increases the evaluation value as said distance from said position distance calculation means becomes larger, and decreases the evaluation value as said distance becomes smaller it is also preferable to cause a computer to function in this way.

[0012] According to another embodiment of the program of the present invention, for the evaluation step, with pseudo movement trajectories of time-series position coordinates and point identifiers used as target data, the position coordinate input layer receives time-series position coordinates of the pseudo movement trajectory, the point identifier input layer receives time-series point identifiers of the pseudo movement trajectory, the position coordinate output layer outputs time-series position coordinates estimated for the pseudo movement trajectory, the point identifier output layer outputs time-series point identifiers estimated for the pseudo movement trajectory, position distance calculation means for calculating a distance between the time-series position coordinates output from the position coordinate output layer and the time-series position coordinates of the pseudo movement trajectory; and point difference calculation means for calculating an error between the time-series point identifiers output from the point identifier output layer and the time-series point identifiers of the pseudo movement trajectory; and evaluation means for calculating an evaluation value based on said distance from said position distance calculation means and said error from said point difference calculation means it is also preferable to cause a computer to function as said means.

[0013] According to other embodiments of the program of the present invention, The evaluation method is, From the standpoint of anonymity, the longer the distance calculated by the position distance calculation means, the higher the evaluation value; the shorter the distance, the lower the evaluation value; From the standpoint of validity, the smaller the error in the point difference calculation method, the higher the evaluation value; and the larger the error, the lower the evaluation value. It is also preferable to make the computer function in this way.

[0014] According to other embodiments of the program of the present invention, For the evaluation phase, the target data will be the pseudo-movement trajectory of time-series position coordinates and point identifiers. The position coordinate input layer receives the time-series position coordinates of the pseudo-movement trajectory, The location identifier input layer receives location identifiers in the time series of the pseudo-movement trajectory. The position coordinate output layer outputs the time-series position coordinates estimated for the pseudo-movement trajectory. The location identifier output layer outputs time-series location identifiers estimated for the pseudo-movement trajectory. The travel time output layer outputs the estimated travel time for the pseudo-travel trajectory. A position distance calculation means calculates the distance between the time-series position coordinates output from the position coordinate output layer and the time-series position coordinates of the pseudo-movement trajectory. A location difference calculation means calculates the error between the time-series location identifiers output from the location identifier output layer and the time-series location identifiers of the pseudo-movement trajectory. A travel time difference calculation means calculates the time difference between the travel time output from the travel time output layer and the travel time of the simulated travel trajectory, An evaluation means that calculates an evaluation value based on the distance calculated by the position distance calculation means, the error calculated by the point difference calculation means, and the time difference calculated by the travel time difference calculation means. It is also preferable to make the computer function in this way.

[0015] According to other embodiments of the program of the present invention, The evaluation method is, From the perspective of anonymity, the longer the distance calculated by the position distance calculation means, the higher the evaluation value, and the shorter the distance, the lower the evaluation value. From the standpoint of validity, the smaller the error in the point difference calculation means, the higher the evaluation value; and the larger the error, the lower the evaluation value. From the standpoint of validity, the shorter the time difference calculated by the travel time difference calculation method, the higher the evaluation value; and the longer the time difference, the lower the evaluation value. It is also preferable to make the computer function in this way.

[0016] According to the present invention, an evaluation device for evaluating movement trajectories, For the training phase, the measured movement trajectory of time-series position coordinates and point identifiers will be used as training data. A position coordinate input layer that converts the time-series position coordinates of the measured movement trajectory into a position coordinate embedding representation, A location identifier input layer that converts location identifiers from the time series of measured movement trajectories into a location identifier embedded representation, A hidden layer that is fully connected to both the position coordinate input layer and the point identifier input layer, A position coordinate output layer is connected to the hidden layer via a fully connected array to estimate the time-series position coordinates. It has, The position coordinate output layer is trained by backpropagation to output time-series position coordinates in the measured movement trajectory, consisting of the position coordinate output layer, a hidden layer, a point identifier input layer, and a position coordinate input layer. It is characterized by the following:

[0017] According to the present invention, there is an evaluation method for a device that evaluates movement trajectories, The device is For the training phase, the measured movement trajectory of time-series position coordinates and point identifiers will be used as training data. A position coordinate input layer that converts the time-series position coordinates of the measured movement trajectory into a position coordinate embedding representation, A location identifier input layer that converts time-series location identifiers of measured movement trajectories into embedded location identifier representations, A hidden layer that is fully connected to both the position coordinate input layer and the point identifier input layer, A position coordinate output layer is connected to the hidden layer via a fully connected array to estimate the time-series position coordinates. It has, The position coordinate output layer is trained by backpropagation to output time-series position coordinates in the measured movement trajectory, consisting of the position coordinate output layer, a hidden layer, a point identifier input layer, and a position coordinate input layer. It is characterized by being executed in such a way. [Effects of the Invention]

[0018] According to the program, evaluation device, and method of the present invention, a simulated movement trajectory can be evaluated from the viewpoint of anonymity with respect to the measured movement trajectory. [Brief explanation of the drawing]

[0019] [Figure 1] This is a system configuration diagram for the present invention. [Figure 2] This is a first functional configuration diagram of the training stage evaluation device according to the present invention. [Figure 3] This is a first functional configuration diagram of the evaluation device for the evaluation stage in the present invention. [Figure 4] This is a second functional configuration diagram of the training stage evaluation device according to the present invention. [Figure 5] This is a second functional configuration diagram of the evaluation device for the evaluation stage in the present invention. [Figure 6] This is a third functional configuration diagram of the training stage evaluation device according to the present invention. [Figure 7] This is a diagram showing the third functional configuration of the evaluation device for the evaluation stage in the present invention. [Modes for carrying out the invention]

[0020] Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings.

[0021] Figure 1 is a system configuration diagram of the present invention.

[0022] As shown in Figure 1, the evaluation device 1 uses a machine learning model and, as part of the training phase, is trained by inputting "measured movement trajectories" as ground truth data. Furthermore, evaluation device 1 uses a machine learning model to input and evaluate "pseudo-movement trajectories" as target data in the <evaluation stage>. Evaluation device 1 evaluates the pseudo-movement trajectories in relation to the measured movement trajectories, at least from the perspective of anonymity and also from the perspective of validity.

[0023] The movement trajectory (measured movement trajectory and simulated movement trajectory) consists of the following data set. Time-series position coordinates Time-series location identifiers

[0024] Time-series position coordinates are the position coordinates over a predetermined travel time (e.g., 24 hours) over a time series (e.g., every 10 minutes). They are a data sequence of discrete position coordinates. Location coordinates can be latitude and longitude, or grid coordinates. Latitude and longitude are a coordinate system that covers the entire globe, expressed in degrees, minutes, and seconds. Grid coordinates, for example, in the case of the UTM grid (Universal Transverse Mercator Grid), divide the Earth into longitude zones of 6 degrees each, and specify the location within each zone using orthogonal coordinates (in meters).

[0025] A time-series point identifier is a time-series (e.g., every 10 minutes) of Points of Interest (PoIs) over a predetermined travel time (e.g., 24 hours). This is also a discrete data sequence of PoIs. A PoI is an IDentifier of the geographical characteristics of a place, and can refer to landmarks, commercial or public facilities, transportation facilities, tourist attractions, etc. These can be obtained from open data.

[0026] <First Training Phase> Figure 2 is a first functional configuration diagram of the training stage evaluation device according to the present invention.

[0027] The evaluation device 1 uses the measured movement trajectory of time-series position coordinates and point identifiers as training data. As shown in Figure 2, the evaluation device 1, in its training phase, includes a position coordinate input layer 11, a location identifier input layer 12, a hidden layer 13, and a position coordinate output layer 141. These functional components are realized by executing a program that enables the computer mounted on the device to function. Furthermore, the processing flow of these functional components can also be understood as a method for evaluating the device.

[0028] The evaluation device 1 is configured, for example, as a fully connected neural network. It consists of three layers: an input layer, a hidden layer, and an output layer, and propagation is unidirectional from the input layer to the output layer. The evaluation device 1 has a position coordinate input layer 11 and a location identifier input layer 12 as input layers.

[0029] [Position coordinate input layer 11] The position coordinate input layer 11 converts the time-series position coordinates of the measured movement trajectory into an "embedded position coordinate representation". The position coordinate input layer 11 has units representing the time series over a predetermined travel time and inputs the position coordinates of the measured travel trajectory in parallel. For example, if the predetermined travel time is 24 hours and there are position coordinates at 144 points in time (= (24 hours × 60 minutes) / 10-minute intervals), the number of input units in the position coordinate input layer 11 will be 144. The position coordinate input layer 11 then outputs a position coordinate embedding representation of a predetermined length via the encoding layer by fully connecting multiple unit layers. These position coordinate embedding representations are then fully connected to the hidden layer 13.

[0030] [Location identifier input layer 12] The location identifier input layer 12 converts the time-series location identifiers of the measured movement trajectory into an "embedded location identifier representation". The location identifier input layer 12 has units for the time series over a predetermined travel time and inputs the location identifiers of the measured travel trajectory in parallel. For example, if the predetermined travel time is 24 hours and there are 24 time points (= (24 hours / 1 hour)) with corresponding position coordinates, the number of units in the location identifier input layer 12 will be 24. The location identifier input layer 12 then outputs a location identifier embedded representation of a predetermined length via the encoding layer by fully connecting multiple unit layers. These location identifier embedded representations are then fully connected to the hidden layer 13.

[0031] [Hidden Layer 13] The hidden layer 13 is fully connected to both the position coordinate input layer 11 and the location identifier input layer 12. The hidden layer 13 is composed of fully connected layers arranged in a graph. Each layer has multiple units, and during the training phase, it calculates the parameters of a function that connects the units from the forward layer to the units in the backward layer. The hidden layer 13 then outputs in parallel from the same number of units as the position coordinates of the measured movement trajectory. For example, similar to the position coordinate input layer 11, it outputs embedded representations from 144 units. These embedded representations are fully coupled to the position coordinate output layer 141.

[0032] [Position coordinate output layer 141] The position coordinate output layer 141 is fully connected to the hidden layer 13 to estimate the time-series position coordinates. The time-series position coordinates output are similar to those of the position coordinate input layer 11. For example, if the predetermined travel time is 24 hours and there are position coordinates at 144 points in time, the number of units in the position coordinate output layer 141 will be 144.

[0033] Then, the position coordinate output layer 141 trains itself, the hidden layer 13, the point identifier input layer 12, and the position coordinate input layer 11 by backpropagation in order to output the time-series position coordinates of the measured movement trajectory. That is, it is trained by backpropagation so that the loss between the time-series position coordinates estimated by itself and the time-series position coordinates of the measured movement trajectory as the target data is minimized (the time-series position coordinates match).

[0034] <First evaluation stage> Figure 3 is a diagram showing the first functional configuration of the evaluation device for the evaluation stage in the present invention.

[0035] According to Figure 3, the evaluation device 1, compared to Figure 2, further includes a position distance calculation unit 142 and an evaluation unit 17 as part of the evaluation stage. For the evaluation stage, the evaluation device 1 inputs a simulated movement trajectory of time-series position coordinates and point identifiers as target data and evaluates it from the perspective of anonymity. The simulated movement trajectory, like the measured movement trajectory, consists of time-series position coordinates and point identifiers at an arbitrary movement time.

[0036] The position coordinate input layer 11 receives the time-series position coordinates of the pseudo-movement trajectory as input. Furthermore, the location identifier input layer 12 receives location identifiers in the time series of the pseudo-movement trajectory. As a result, the position coordinate output layer 141 outputs the time-series position coordinates estimated for the pseudo-movement trajectory. These time-series position coordinates are reconstructed from the pseudo-movement trajectory using a neural network trained on the actual movement trajectory.

[0037] [Position distance calculation unit 142] The position distance calculation unit 142 calculates the distance between the time-series position coordinates output from the position coordinate output layer 141 and the time-series position coordinates of the pseudo-movement trajectory. This "distance" represents the error (similarity, loss) between the two time-series position coordinates.

[0038] The distance error may be calculated, for example, by dynamic time warping (DTW) or cross-entropy loss.

[0039] Dynamic time warping is an algorithm used to measure the similarity between two time series. It involves arranging the two time series in a grid and calculating the distance between each point. Next, it detects the path (warping path) that minimizes the distance from the start point to the end point. The distance calculated along the warping path represents the similarity between the two time series. This method can calculate similarity even between time series with different time lengths or temporal lags.

[0040] Cross-entropy loss calculates the difference between two time series. For example, the Kullback-Leibler divergence (KL divergence) can be used. P: Time series position coordinates as estimated data Q: Time-series position coordinates of the pseudo-movement trajectory as the target data KL(P||Q)=Σ i=1 N P(xi)·log(P(xi) / Q(xi)) KL divergence represents the amount of information loss that occurs when Q is approximated using P. Other embodiments may include JS divergence or Wasserstein distance.

[0041] [Evaluation Section 17] The evaluation unit 17 calculates an evaluation value based on the distance calculated by the position distance calculation unit 142.

[0042] Based on anonymity considerations, the evaluation results are determined as follows: Longer distance between coordinates (lower similarity) -> higher anonymity -> higher evaluation score Shorter distance between coordinates (higher similarity) -> lower anonymity -> lower evaluation score.

[0043] In other words, "a long distance in positional coordinates" means that the movement trajectory of similar estimated data has not been reconstructed for the pseudo-movement trajectory of the target data (i.e., different movement trajectories have been estimated). This is desirable from the standpoint of anonymity (privacy protection) of movement trajectories. On the other hand, "short distance in positional coordinates" means that the movement trajectories of similar estimated data are being reconstructed for the pseudo-movement trajectory of the target data (i.e., similar movement trajectories are being estimated). This is undesirable from the standpoint of anonymity of movement trajectories.

[0044] <Second Training Phase> Figure 4 is a second functional configuration diagram of the training stage evaluation device according to the present invention.

[0045] According to Figure 4, the evaluation device 1, in the training phase, further includes a location identifier output layer 151 compared to Figure 2.

[0046] [Location identifier output layer 151] The location identifier output layer 151 is fully connected to the position coordinate output layer 141 to estimate the time-series location identifier. Similar to the location identifier input layer 12, the location identifier output layer 151, for example, if a predetermined travel time is 24 hours and there are location identifiers for 24 time points, will have 24 input units in the location identifier input layer 12.

[0047] The location identifier output layer 151 then trains the location identifier output layer, position coordinate output layer, hidden layer, location identifier input layer, and position coordinate input layer by backpropagation in order to output time-series location identifiers in the measured movement trajectory. In other words, it is trained by backpropagation so that the loss between the time-series location identifiers it estimates and the time-series position coordinates of the measured movement trajectory as the target data is minimized (the time-series location identifiers match).

[0048] <Second evaluation stage> Figure 5 is a second functional configuration diagram of the evaluation device for the evaluation stage in the present invention.

[0049] According to Figure 5, the evaluation device 1, compared to Figure 4, further includes a position distance calculation unit 142, a point difference calculation unit 152, and an evaluation unit 17 for the evaluation stage. For the evaluation stage, the evaluation device 1 inputs the time-series position coordinates and the simulated movement trajectory of the point identifier as target data.

[0050] As shown in Figure 3 above, the evaluation unit 17 evaluates only from the perspective of "anonymity" in the simulated movement trajectory. In other words, it will give a high evaluation value to a random simulated movement trajectory that is completely different from the measured movement trajectory. In contrast, Figure 5 shows that the evaluation from the perspective of "validity" is added to the evaluation from the perspective of "anonymity" in Figure 3.

[0051] The position coordinate input layer 11 receives the time-series position coordinates of the pseudo-movement trajectory as input. Furthermore, the location identifier input layer 12 receives location identifiers in the time series of the pseudo-movement trajectory. In response, the position coordinate output layer 141 outputs the time-series position coordinates estimated for the pseudo-movement trajectory. As a result, the location identifier output layer 151 outputs a time-series of location identifiers estimated for the pseudo-movement trajectory. These time-series location identifiers are reconstructed from the pseudo-movement trajectory using a neural network trained on the actual movement trajectory.

[0052] [Position distance calculation unit 142] This is exactly the same as the position distance calculation unit 142 mentioned above.

[0053] [Location difference calculation unit 152] The location difference calculation unit 152 calculates the error between the time-series location identifiers output from the location identifier output layer 151 and the time-series location identifiers of the pseudo-movement trajectory.

[0054] The difference in location identifiers may be calculated, for example, by cross-entropy loss. As mentioned earlier, cross-entropy loss calculates the difference between two time series. For example, the Kullback-Leibler divergence can be used. P: Time series location identifier as estimated data Q: Point identifiers for the time series of pseudo-movement trajectories as target data. KL(P||Q)=Σ i=1 N P(xi)·log(P(xi) / Q(xi)) KL divergence represents the amount of information loss that occurs when Q is approximated using P. Other embodiments may include JS divergence or Wasserstein distance.

[0055] [Evaluation Section 17] The evaluation unit 17 calculates an evaluation value based on the distance calculated by the position distance calculation unit 142 and the error calculated by the point difference calculation unit 152.

[0056] Based on the perspectives of anonymity and validity, the evaluation values ​​are determined as follows: (From the perspective of anonymity) Longer distance between coordinates (lower similarity) -> higher anonymity -> higher evaluation score Shorter distance between coordinates (higher similarity) -> lower anonymity -> lower evaluation score. (From the perspective of validity) Small error in location identifiers (high similarity) -> high validity -> higher evaluation score. Large error in location identifiers (low similarity) -> Low validity -> Lower evaluation score.

[0057] "A small error in location identifiers" means that the movement trajectories of similar estimated data are reconstructed based on the pseudo-movement trajectory of the target data (i.e., similar movement trajectories are estimated). This is desirable from the standpoint of validity of the movement trajectory. On the other hand, "large error in location identifiers" means that the movement trajectories of similar estimated data have not been reconstructed for the pseudo-movement trajectories of the target data (i.e., different movement trajectories have been estimated). This is undesirable from the standpoint of validity of the movement trajectories.

[0058] <Third Training Stage> Figure 6 is a diagram showing the third functional configuration of the training stage evaluation device according to the present invention.

[0059] According to Figure 6, the evaluation device 1, as a <training stage>, further includes a travel time output layer 161 compared to Figure 4.

[0060] [Travel time output layer 161] The travel time output layer 161 is fully connected to the location identifier output layer 151 to estimate the travel time (trip time between Origin and Destination). The travel time output layer 161 has one input unit for a predetermined travel time.

[0061] Then, the travel time output layer 161 trains itself, the location identifier output layer 151, the position coordinate output layer 141, the hidden layer 13, the location identifier input layer 12, and the position coordinate input layer 11 by backpropagation in order to output the travel time in the measured travel trajectory. The travel time output layer 161 is trained by backpropagation so that the loss between the travel time it estimates and the travel time of the measured travel trajectory as the target data is minimized (the travel times match).

[0062] <Third evaluation stage> Figure 7 is a diagram showing the third functional configuration of the evaluation device for the evaluation stage in the present invention.

[0063] According to Figure 7, the evaluation device 1, in the <evaluation stage>, further includes a position distance calculation unit 142, a point difference calculation unit 152, a travel time difference calculation unit 162, and an evaluation unit 17, compared to Figure 6. For the evaluation stage, the evaluation device 1 inputs the time-series position coordinates and the simulated movement trajectory of the point identifier as target data.

[0064] As shown in Figure 7, similar to Figure 5, the evaluation from the perspective of "anonymity" is combined with the evaluation from the perspective of "validity."

[0065] The position coordinate input layer 11 receives the time-series position coordinates of the pseudo-movement trajectory as input. Furthermore, the location identifier input layer 12 receives location identifiers in the time series of the pseudo-movement trajectory. As a result, the position coordinate output layer 141 outputs the time-series position coordinates estimated for the pseudo-movement trajectory. Furthermore, the location identifier output layer 151 outputs a time-series of location identifiers estimated for the pseudo-movement trajectory. As a result, the travel time output layer 161 outputs the estimated travel time for the pseudo-travel trajectory. This travel time is estimated from the pseudo-travel trajectory using a neural network trained on the actual travel trajectory.

[0066] [Position distance calculation unit 142] This is exactly the same as the position distance calculation unit 142 mentioned above.

[0067] [Location difference calculation unit 152] This is exactly the same as the point difference calculation unit 152 mentioned above.

[0068] [Travel time difference calculation unit 162] The travel time difference calculation unit 162 calculates the time difference between the travel time estimated by the travel time output layer 161 and the travel time of the simulated travel trajectory. The time difference may be, for example, the Mean Squared Error (MSE).

[0069] [Evaluation Section 17] The evaluation unit 17 calculates an evaluation value based on the distance calculated by the position distance calculation unit 142, the error calculated by the point difference calculation unit 152, and the time difference calculated by the travel time difference calculation unit 162.

[0070] Based on the perspectives of anonymity and validity, the evaluation values ​​are determined as follows: (From the perspective of anonymity) Longer distance between coordinates (lower similarity) -> higher anonymity -> higher evaluation score Shorter distance between coordinates (higher similarity) -> lower anonymity -> lower evaluation score. (From the perspective of validity) Small error in location identifiers (high similarity) -> high validity -> higher evaluation score. Large error in location identifiers (low similarity) -> Low validity -> Lower evaluation score. (From the perspective of validity) Short travel time difference (high similarity) -> high validity -> higher evaluation score Long travel time difference (low similarity) -> low validity -> lower evaluation score.

[0071] "A short difference in travel time" means that the travel trajectories of similar estimated data are reconstructed based on the pseudo-travel trajectory of the target data (i.e., similar travel times are estimated). This is desirable from the standpoint of validating the travel trajectory. On the other hand, a "long difference in travel time" means that the travel trajectories of similar estimated data have not been reconstructed in relation to the pseudo-travel trajectory of the target data (i.e., different travel times have been estimated). This is undesirable from the standpoint of validity of the travel trajectory.

[0072] Here, the evaluation unit 17 may control the three losses used to derive the evaluation value using hyperparameters α, β, and γ, and calculate the combined integrated loss. Integrated loss = α loss1 + β loss2 + γ loss3 loss1: Distance from position coordinates loss2: Error in location identifier Loss 3: Travel time difference

[0073] As described in detail above, the program, evaluation device, and method of the present invention make it possible to evaluate simulated movement trajectories from the viewpoint of anonymity with respect to measured movement trajectories. Furthermore, according to the present invention, it is also possible to comprehensively evaluate the simulated movement trajectory, including from the viewpoint of its validity with respect to the measured movement trajectory.

[0074] Furthermore, this will enable, for example, the evaluation of simulated movement trajectories usable across a variety of services, thereby contributing to Goal 9 of the United Nations-led Sustainable Development Goals (SDGs): "Build resilient infrastructure, promote sustainable industrialization and foster innovation."

[0075] Various changes, modifications, and omissions to the scope of the technical concept and viewpoint of the present invention can be readily made by those skilled in the art with respect to the various embodiments of the present invention described above. The above description is merely illustrative and is not intended to limit the present invention in any way. The present invention is limited only to what is limited by the claims and their equivalents. [Explanation of Symbols]

[0076] 1. Evaluation device 11 Position Coordinate Input Layer 12. Location Identifier Input Layer 13 Hidden Layers 141 Position Coordinate Output Layer 142 Position distance calculation section 151 Location Identifier Output Layer 152 Point Difference Calculation Unit 161 Transit Time Output Layer 162 Travel Time Difference Calculation Unit 17 Evaluation Department 2. Database of measured movement trajectories 3. Pseudo-movement trajectory database

Claims

1. A program that enables the operation of a computer mounted on a device that evaluates movement trajectories, For the training phase, the measured movement trajectory of time-series position coordinates and point identifiers will be used as training data. A position coordinate input layer that converts the time-series position coordinates of the measured movement trajectory into a position coordinate embedding representation, A location identifier input layer that converts time-series location identifiers of measured movement trajectories into embedded location identifier representations, A hidden layer that is fully connected to both the position coordinate input layer and the point identifier input layer, A position coordinate output layer is connected to the hidden layer via a fully connected array to estimate the time-series position coordinates. and make it work The position coordinate output layer is trained by backpropagation to output time-series position coordinates in the measured movement trajectory, consisting of the position coordinate output layer, a hidden layer, a point identifier input layer, and a position coordinate input layer. A program characterized by causing a computer to function in a certain way.

2. Regarding the training phase, A location identifier output layer is connected to the position coordinate output layer via a fully connected interface, and estimates the time-series location identifier. And make it work even better. The location identifier output layer is trained by backpropagation to output time-series location identifiers in the measured movement trajectory, along with the location coordinate output layer, the hidden layer, the location identifier input layer, and the location coordinate input layer. The program according to claim 1, characterized in that it causes the computer to function in such a way.

3. Regarding the training phase, A travel time output layer is connected to the location identifier output layer in a fully connected manner to estimate travel time. And make it work even better. The movement time output layer is trained by backpropagation to output the movement time in the measured movement trajectory. This training includes the movement time output layer, the point identifier output layer, the position coordinate output layer, the hidden layer, the point identifier input layer, and the position coordinate input layer. The program according to claim 2, characterized in that it causes the computer to function in such a way.

4. For the evaluation phase, the target data will be the pseudo-movement trajectory of time-series position coordinates and point identifiers. The position coordinate input layer receives the time-series position coordinates of the pseudo-movement trajectory, The location identifier input layer receives location identifiers in the time series of the pseudo-movement trajectory. The position coordinate output layer outputs the time-series position coordinates estimated for the pseudo-movement trajectory. A position distance calculation means calculates the distance between the time-series position coordinates output from the position coordinate output layer and the time-series position coordinates of the pseudo-movement trajectory. An evaluation means that calculates an evaluation value based on the distance calculated by the position distance calculation means. The program according to claim 1, characterized in that it causes a computer to function.

5. The evaluation method is, From the perspective of anonymity, the longer the distance calculated by the positional distance calculation method, the higher the evaluation value, and the shorter the distance, the lower the evaluation value. The program according to claim 4, characterized in that it causes the computer to function in such a way.

6. For the evaluation phase, the target data will be the pseudo-movement trajectory of time-series position coordinates and point identifiers. The position coordinate input layer receives the time-series position coordinates of the pseudo-movement trajectory, The location identifier input layer receives location identifiers in the time series of the pseudo-movement trajectory. The position coordinate output layer outputs the time-series position coordinates estimated for the pseudo-movement trajectory. The location identifier output layer outputs time-series location identifiers estimated for the pseudo-movement trajectory. A position distance calculation means calculates the distance between the time-series position coordinates output from the position coordinate output layer and the time-series position coordinates of the pseudo-movement trajectory. A location difference calculation means calculates the error between the time-series location identifiers output from the location identifier output layer and the time-series location identifiers of the pseudo-movement trajectory. An evaluation means that calculates an evaluation value based on the distance calculated by the position distance calculation means and the error calculated by the point difference calculation means. The program according to claim 2, characterized in that it causes a computer to function.

7. The evaluation method is, From the standpoint of anonymity, the longer the distance calculated by the position distance calculation means, the higher the evaluation value; the shorter the distance, the lower the evaluation value; From the standpoint of validity, the smaller the error in the point difference calculation method, the higher the evaluation value; and the larger the error, the lower the evaluation value. The program according to claim 6, characterized in that it causes the computer to function in such a way.

8. For the evaluation phase, the target data will be the pseudo-movement trajectory of time-series position coordinates and point identifiers. The position coordinate input layer receives the time-series position coordinates of the pseudo-movement trajectory, The location identifier input layer receives location identifiers in the time series of the pseudo-movement trajectory. The position coordinate output layer outputs the time-series position coordinates estimated for the pseudo-movement trajectory. The location identifier output layer outputs time-series location identifiers estimated for the pseudo-movement trajectory. The travel time output layer outputs the estimated travel time for the pseudo-travel trajectory. A position distance calculation means calculates the distance between the time-series position coordinates output from the position coordinate output layer and the time-series position coordinates of the pseudo-movement trajectory. A location difference calculation means calculates the error between the time-series location identifiers output from the location identifier output layer and the time-series location identifiers of the pseudo-movement trajectory. A travel time difference calculation means calculates the time difference between the travel time output from the travel time output layer and the travel time of the simulated travel trajectory, An evaluation means that calculates an evaluation value based on the distance calculated by the position distance calculation means, the error calculated by the point difference calculation means, and the time difference calculated by the travel time difference calculation means. The program according to claim 3, characterized in that it causes a computer to function.

9. The evaluation method is, From the perspective of anonymity, the longer the distance calculated by the position distance calculation means, the higher the evaluation value, and the shorter the distance, the lower the evaluation value. From the standpoint of validity, the smaller the error in the point difference calculation means, the higher the evaluation value; and the larger the error, the lower the evaluation value. From the standpoint of validity, the shorter the time difference calculated by the travel time difference calculation method, the higher the evaluation value; and the longer the time difference, the lower the evaluation value. The program according to claim 8, characterized in that it causes the computer to function in such a way.

10. An evaluation device for evaluating movement trajectories, For the training phase, the measured movement trajectory of time-series position coordinates and point identifiers will be used as training data. A position coordinate input layer that converts the time-series position coordinates of the measured movement trajectory into a position coordinate embedding representation, A location identifier input layer that converts time-series location identifiers of measured movement trajectories into embedded location identifier representations, A hidden layer that is fully connected to both the position coordinate input layer and the point identifier input layer, A position coordinate output layer is connected to the hidden layer via a fully connected array to estimate the time-series position coordinates. It has, The position coordinate output layer is trained by backpropagation to output time-series position coordinates in the measured movement trajectory, consisting of the position coordinate output layer, a hidden layer, a point identifier input layer, and a position coordinate input layer. An evaluation device characterized by the following features.

11. A method for evaluating a device that evaluates movement trajectories, The device is For the training phase, the measured movement trajectory of time-series position coordinates and point identifiers will be used as training data. A position coordinate input layer that converts the time-series position coordinates of the measured movement trajectory into a position coordinate embedding representation, A location identifier input layer that converts time-series location identifiers of measured movement trajectories into embedded location identifier representations, A hidden layer that is fully connected to both the position coordinate input layer and the point identifier input layer, A position coordinate output layer is connected to the hidden layer via a fully connected array to estimate the time-series position coordinates. It has, The position coordinate output layer is trained by backpropagation to output time-series position coordinates in the measured movement trajectory, consisting of the position coordinate output layer, a hidden layer, a point identifier input layer, and a position coordinate input layer. A method for evaluating an apparatus, characterized by being performed in such a manner.

Images