Selection device, selection method, and program

The selection device and method address the lack of explainability in VRP solvers by selecting edges based on user attributes and graph information, providing explanations that align with user preferences and enhance route understanding.

WO2026239576A1PCT designated stage Publication Date: 2026-11-19NT T INC
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
PCT/JP2025/017803
Authority / Receiving Office
WO · WO
Patent Type
Applications
Current Assignee / Owner
Filing Date
2025-05-16
Publication Date
2026-11-19

AI Technical Summary

Technical Problem

Existing VRP solvers lack the ability to provide explanations for generated routes, making it difficult for users to understand why certain edges were chosen over others, especially when the number of edges is large or the number of counterfactual virtual edges is enormous.

Method used

A selection device and method that selects an appropriate set of edges for explanation based on user attributes and graph information, using a recommendation model to calculate recommendation scores and identify the real edge that least matches user criteria and the counterfactual virtual edge that best matches user criteria.

Benefits of technology

Enables the generation of explanations for VRP routes that align with user preferences, simplifying the selection process and improving user understanding and satisfaction with the route.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure JP2025017803_19112026_PF_FP_ABST
    Figure JP2025017803_19112026_PF_FP_ABST
Patent Text Reader

Abstract

A selection device according to one aspect disclosed herein selects: a first edge included in a route on a predetermined graph; and a second edge having the same start point as the first edge and a different end point from the first edge. The selection device comprises a selection unit for selecting the first edge from among edges constituting the route and the second edge from among edges having the same start point as the first edge and a different end point from the first edge on the basis of user attribute information indicating a user attribute, the route, and information about each point included in the graph.
Need to check novelty before this filing date? Find Prior Art

Description

Selection device, selection method, and program

[0001] This disclosure relates to a selection device, a selection method, and a program.

[0002] In many cases, when considering travel between points on a route (which may also be called a "path") generated as an optimal solution to problems such as the Vehicle Routing Problem (VRP), it is necessary to explain "why that particular route is optimal."

[0003] For this reason, a technique is known that can generate an explanation for the question, "Why was the actual edge selected instead of the counterfactual edge?" for a pair of edges on a route (hereinafter also called a "real edge") and an edge that has the same starting point as that edge but a different destination (hereinafter also called a "counterfactual virtual edge") (Non-Patent Literature 1).

[0004] D. Kikuta et al., RouteExplainer: An Explanation Framework for Vehicle Routing Problem, PAKDD 2024.

[0005] However, in the technology described in Non-Patent Document 1, the user must specify the pair of real edges and counterreal virtual edges to be explained. For this reason, if, for example, the number of edges constituting a route is large, or if the number of counterreal virtual edges relative to real edges is enormous, it may be difficult to select the real edges and counterreal virtual edges to be explained.

[0006] This disclosure is made in view of the above points and aims to provide a technology that can select an appropriate set of edges to be described.

[0007] A selection device according to one aspect of the present disclosure is a selection device that selects a first edge included in a path on a predetermined graph, and a second edge that shares the same start point as the first edge and has a different end point, the selection device comprising a selection unit that selects the first edge from edges constituting the path, and selects the second edge from edges that share the same start point as the first edge and have different end points, based on user attribute information representing a user's attributes, the path, and information related to each point included in the graph.

[0008] An appropriate set of edges to be explained can be selected.

[0009] It is a diagram illustrating an example of conventional technology. It is a diagram illustrating an example of the proposed technology. It is a diagram illustrating an example of the hardware configuration of an explanation device according to an embodiment. It is a diagram illustrating an example of the functional configuration of the explanation device according to the embodiment. It is a flowchart illustrating an example of explanation processing according to the embodiment.

[0010] Hereinafter, one embodiment of the present invention will be described in detail with reference to the drawings.

[0011] <Background> There are many cases where an explanation of "why that movement is optimal" is required for movement between certain points on a route generated as an optimal solution for a vehicle routing problem (VRP) or the like. For example, in the generation of dispatch routes for power supply vehicles when a large-scale power outage occurs, accountability for the dispatch route arises. Similarly, for example, in the generation of dispatch routes for mobile base stations when a large-scale failure occurs in a wireless communication network, accountability for the dispatch route arises. Further, for example, when a sightseeing route or the like is generated and proposed interactively, explaining "why the movement is optimal" regarding movement along the proposed route leads to improvement of the user's satisfaction and understanding, and it is considered that the user can easily modify the route by taking the explanation into consideration.

[0012] On the other hand, since existing VRP solvers do not take explainability into account, a VRP solver alone cannot provide an explanation for the generated route.

[0013] <Prior Art> As a technique that can provide an explanation for routes generated by a VRP solver or the like (hereinafter also referred to as "actual routes"), the technique described in Non-Patent Document 1 is known. Below, an outline of the technique described in Non-Patent Document 1 will be explained. In the following, as an example, the case in which the actual route is a circular route will be explained.

[0014] Figure 1 shows an example of the conventional technology. As shown in Figure 1, (1-1) a real route is given. In the example shown in Figure 1, the circulating route 1→2→3→4→6→7→5→1 is given as the real route. At this time, (1-2) the user wants to know "why the movement 3→4 was selected instead of 3→5". In this case, edge 3-4, which represents the movement 3→4, is set as the real edge to be explained, and edge 3-5, which represents the movement 3→5, is set as the counterfactual virtual edge to be explained. Here, a real edge is an edge included in the real route, and a counterfactual virtual edge is an edge that has the same starting point (start point) as the real edge in question, but a different destination (end point). Note that a counterfactual virtual edge is an edge that is not included in the real route, but could have been included in the real route.

[0015] Next, (1-3) a counterfactual route is generated using a VRP solver or the like, which is a route that follows the counterfactual edge. A counterfactual route is a route that includes a counterfactual edge, and is generated with the edges prior to the counterfactual edge being the same as the real route. In the example shown in Figure 1, the cyclic route 1→2→3→5→6→7→1 is generated as a counterfactual route.

[0016] Next, (1-4) using a classification model that classifies edges into one of several predetermined classes, we classify the edges included in the real route and the edges included in the counterreal virtual route, and assign a label to each edge that represents the class to which it belongs. In the example shown in Figure 1, the label y is assigned to edge 1-2 included in the real route. 1 fact Label y for edge 2-3 2 fact Label y for edge 3-4 3fact , label y for edge 4-6 4 fact , label y for edge 6-7 5 fact , label y for edge 7-5 6 fact , label y for edge 5-1 7 fact are respectively assigned. Similarly, label y for edge 1-2 included in the counterfactual route 1 Q , label y for edge 2-3 2 Q , label y for edge 3-5 3 Q , label y for edge 5-6 4 Q , label y for edge 6-7 5 Q , label y for edge 7-1 6 Q are respectively assigned. Note that, as an example, when the VRP is a traveling salesman problem with time windows (TSPTW), a model that assigns labels that can take values representing visit reasons (e.g., "because the travel distance can be shortened", "to comply with time constraints", etc.) can be used as the classification model.

[0017] Next, in (1-5), the influence degree of the target real edge to be explained is calculated using the edges included in the real route and their labels. Similarly, the influence degree of the target counterfactual edge to be explained is calculated using the edges included in the counterfactual route and their labels. Then, the influence degree of the target real edge to be explained is compared with the influence degree of the target counterfactual edge to be explained. Note that, as an example, when the VRP is a traveling salesman problem with time windows, as the influence degree of an edge, at least one of the length of the edge sequence up to that edge (e.g., the sum of the Euclidean distances of each edge included in the edge sequence), the proportion of each value of the label assigned to each edge included in the edge sequence, the proportion of points that can be visited on the route represented by the edge sequence, and the like is used.

[0018] (1-6) Using the comparison results between the influence of the actual edge being explained and the influence of the counterfactual virtual edge being explained (e.g., the difference between the influence of the actual edge being explained and the influence of the counterfactual virtual edge being explained), and a natural language sentence template, an explanatory sentence is generated and output in response to the question, "Why was the movement from 3 to 4 chosen instead of 3 to 5?". Note that in (1-6), the explanatory sentence may be generated using, for example, a Large Language Model (LLM).

[0019] Therefore, by referring to the final outputted explanation, the user can obtain an explanation for the question, "Why was the movement 3→4 selected instead of 3→5?"

[0020] <Problems with the Conventional Technology> As shown in (1-2) above, the technology described in Non-Patent Document 1 requires the user to specify the pair of real edges and counterreal virtual edges to be explained. For this reason, for example, if there are many edges that make up a real route, it may be difficult to select the real edges to be explained. Furthermore, even if a real edge to be explained can be selected, if the number of counterreal virtual edges for that real edge is enormous, it may be difficult to select the counterreal virtual edges to be explained.

[0021] For example, if a real route traversing N points is generated from a complete graph consisting of N points, the number of choices for real edges and counterreal virtual edges to be explained is O(N). 2 Therefore, the larger N becomes, the more options there are.

[0022] <Proposed Technology> Below, we propose a technology for selecting real edges and counterreal virtual edges to be explained in order to solve the problems of the conventional technology described above. The proposed technology focuses on the fact that "when a user selects edges that make up a route, they select the edges based on some kind of criteria," and selects the real edge that least matches the user's criteria from among the real edges that make up the real route as the "real edge to be explained," and the counterreal virtual edge that best matches the user's criteria from among the counterreal virtual edges for that real edge as the "counterreal virtual edge to be explained." This is because it is thought that the user seeks an explanation for the real edge that least matches their criteria and selects the counterreal virtual edge that best matches their criteria as the comparison target. Examples of criteria that users use to select edges include preference, intuition, and rationality (e.g., rationality regarding travel costs such as time and fees).

[0023] As an example, let's assume that the route is a "tourist route," and that "when a user selects the edges that make up the route, they select the edges based on their own preferences." Therefore, in this case, among the actual edges that make up the actual route, the actual edge that least matches the user's preferences is selected as the "actual edge to be explained," and among the counterfactual virtual edges for that actual edge, the counterfactual virtual edge that best matches the user's preferences is selected as the "counterfactual virtual edge to be explained."

[0024] Figure 2 shows an example of the proposed technology. As shown in Figure 2, (2-1) a real route generated by a VRP solver or the like is provided. (2-2) A model called a recommendation model is used to calculate a probability called the recommendation score for each real edge included in the real route. The recommendation model is a model that takes user attribute information representing the user's attributes (e.g., gender, age, hobbies, preferences, favorite atmosphere, etc.), information about each point (node) (e.g., coordinates of each point, type of facility represented by each point, etc.), and a visit history representing the history of visited locations as input, and outputs a probability representing the recommendation score for the next place to visit. Since an edge represents movement from one point to another, the recommendation score can be calculated for each real edge included in the real route by using the recommendation model. Here, the recommendation model is realized by training a machine learning model such as a neural network using a learning method such as supervised learning. In the example shown in Figure 2, the recommendation scores calculated are 0.7 for actual edge 1-2, 0.8 for actual edge 2-3, 0.2 for actual edge 3-4, 0.6 for actual edge 4-6, 0.9 for actual edge 6-7, and 0.5 for actual edge 7-5. Since a patrol route is assumed as the actual route, it is not necessary to calculate a recommendation score for actual edge 5-1.

[0025] Next, a set of real edges and counterreal virtual edges to be explained is selected using either method 1 or 2 for selecting the edge set to be explained.

[0026] In Method 1 for Selecting Edge Sets to Explain, first, (2-3) the real edge with the lowest recommendation score among the real edges included in the real route is selected as the "real edge to explain". In the example shown in Figure 2, real edges 3-4 are selected as the "real edge to explain". Next, (2-4) the recommendation score is calculated for the counterfactual edges relative to the real edge to explain using the recommendation model. These counterfactual edges are candidates for the counterfactual edges to explain. In the example shown in Figure 2, a recommendation score of 0.05 is calculated for counterfactual edge 3-6, a recommendation score of 0.05 is calculated for counterfactual edge 3-7, and a recommendation score of 0.7 is calculated for counterfactual edge 3-5. Then, (2-5) the counterfactual edge with the highest recommendation score among the candidate counterfactual edges is selected as the "counterfactual edge to explain".

[0027] In method 2 for selecting the edge pairs to be explained, first, (2-3') the actual edges for which recommendation scores have been calculated are designated as candidates for the actual edges to be explained. In the example shown in Figure 2, actual edges 1-2, 2-3, 3-4, 4-6, 6-7, and 7-5 are candidates for the actual edges to be explained. (2-4') Using the recommendation model, the recommendation scores for counteracting virtual edges for each of the candidate actual edges to be explained are calculated. In the example shown in Figure 2, the recommendation scores for counteracting virtual edges 1-4, 1-3, 1-5, 1-6, and 1-7 for actual edge 1-2 are calculated as 0.5, 0.45, 0.3, 0.4, and 0.3, respectively. Similarly, recommendation scores are calculated for the counteracting virtual edges for actual edges 2-3, 3-4, 4-6, 6-7, and 7-5. Then, (2-5') the real edge and counterfactual edge with the largest recommendation score for a counterfactual edge, obtained by subtracting the recommendation score for the corresponding real edge (hereinafter also referred to as the "recommendation score difference"), are selected as the "real edge to be explained" and the "counterfactual edge to be explained".

[0028] As a result, the real edge that least matches the user's preferences is selected as the "real edge to be explained," and among the counterfactual edges relative to that real edge, the counterfactual virtual edge that least matches the user's preferences is selected as the "counterfactual virtual edge to be explained." Therefore, an appropriate pair of real and counterfactual edges to be explained is selected, eliminating the need for the user to select the real and counterfactual edges to be explained.

[0029] <Details of the proposed technology> The details of the proposed technology described above will be explained below.

[0030] ≪Definition of Notation≫ A VRP instance is defined as a graph G = (V, E, X). Here, V is a set of points (nodes), and each node is v i Let ∈ V, where i = 1, ..., N. Note that N = |V|. Hereafter, node v i This is also called "node i" or "point i," and i is also called "node ID" or "node number." Furthermore, E is a set of directed edges, and the directed edges from node i to node j are e i,j Let ∈ E, where i ≠ j. That is, graph G is a complete graph. Furthermore, X is a set of information about each node (e.g., the coordinates of each node, the type of facility each point represents, etc.), and node v i Information about x i Let ∈X.

[0031] The route generated from graph G is shown below.

[0032]

[0033] Here, π t ∈{1, ..., N} is the node ID of the t-th visited node (hereinafter also referred to as the "visited location"), and T is the number of steps (number of moves) in the route. Note that t∈{1, ..., T} will be referred to as the "step number," etc. Hereafter in the text of this specification, subscripts of subscripts will be represented by adding an underscore "_" immediately before them. For example, in the text of this specification, the route r shown in number 1 above will be expressed as r = (π 1 , π 2 , ..., πT ) = (e π_1,π_2 , e π_2,π_3 Let us represent this as , ..., eπ_(T-1), π_T).

[0034] Let user attribute information be 'a'. User attribute information 'a' is represented as vector data that shows various attributes of the user, such as gender, age, occupation, annual income, hobbies, preferences, and preferred atmosphere.

[0035] A VRP instance is information that includes, for example, the coordinates of each location, the starting point, and the destination. However, a VRP instance may include various other types of information depending on the type and nature of the problem. For example, if the VRP is a time-constrained traveling salesman problem, the VRP instance may include the travel time to each edge and the time slots for visiting each location. In addition, for example, if the VRP is a prize-collecting traveling salesman problem (PCTSP), the VRP instance may include the prizes and penalties at each location. Similarly, for example, if the VRP is a capacity-constrained vehicle routing problem (CVRP), the VRP instance may include the capacity of the delivery vehicles and the demand at each location.

[0036] Furthermore, defining a VRP instance as a complete graph is just one example; it may also be defined as an incomplete graph, for example, if certain points are not connected.

[0037] ≪Recommendation Model≫ The recommendation model uses user attribute information a, information X about each location, and visit history π representing the history of visited locations. 1 , ..., π t-1 Taking and as input, it outputs a probability representing the recommendation level of the next visited location. Hereafter, the recommendation model will be referred to as RecModel.

[0038] The recommendation model RecModel uses, for example, user attribute information a, location information X, and visit history π. 1 , ..., π t-1And the correct destination π represents the correct next place to visit. t This is achieved by training a machine learning model, such as a neural network, using training data composed of the above, and employing methods such as supervised learning. In other words, the recommendation model RecModel is composed of, for example, user attribute information a, information about each location X, and visit history π. 1 , ..., π t-1 The input data is π, the correct destination is π. t This is achieved by using this data as training data and training machine learning models such as neural networks using methods such as supervised learning.

[0039] The recommendation model RecModel can be represented, for example, as follows:

[0040]

[0041] Here, p can be expressed, for example, as follows:

[0042]

[0043] However, in the above equation 3, for simplicity, the case of visiting the same location multiple times is excluded, and l ∈ {1, ..., N} \ {π} 1 , ..., π t-1 This is the case. If multiple visits to the same location are allowed, this part becomes l ∈ {1, ..., N}.

[0044] Furthermore, if the output p of the recommended model RecModel shown in equation 2 above is expressed in component form, then equation 1 above can be expressed as follows.

[0045]

[0046] ≪Proposed technology when using method 1 for selecting the edge pair to be explained≫ (2-1) to (2-5) of the proposed technology can be realized as follows.

[0047] First, calculate the following:

[0048]

[0049] However, the following applies:

[0050]

[0051] In other words, the step number corresponding to the real edge with the lowest recommendation level among the real edges included in the real route is calculated (i.e., the step number of the destination visited by the real edge with the lowest recommendation level). Hereafter, in the text of this specification, a hat symbol "^" will be placed directly above a symbol to represent it. For example, in the text of this specification, the left side of the above number 5 will be represented as ^t.

[0052] At this time, the actual edge e being explained fact The following applies:

[0053]

[0054] Note that in the above number 6, the visit history is (π t-1 The input is set to include only the node IDs of the visited locations visited in the previous step. This is to ensure consistency between the VRP solver that generated the actual route and the recommendation model RecModel. More specifically, for example, the VRP solver generates the optimal route from an objective standpoint such as time constraints, while the recommendation model RecModel recommends the optimal route from the standpoint of user attributes. Therefore, the visit history of the actual route generated by the VRP solver may be a visit history that could not occur with the recommendation model RecModel alone. To avoid situations where the input is a visit history that could not occur with the recommendation model RecModel alone, the recommendation model RecModel uses a visit history that includes only the node IDs of the visited locations visited in the previous step as input.

[0055] Next, calculate the following:

[0056]

[0057] However, the following applies:

[0058]

[0059] In other words, the actual edge e being explained fact Among the counterfactual edges for the target, calculate the node number of the destination visited by the counterfactual edge with the highest recommendation score.

[0060] At this point, the counterfactual edge e being explained cf The following applies:

[0061]

[0062] Furthermore, in step 9 above, in order to ensure consistency between the VRP solver and the recommendation model RecModel, the input to the recommendation model RecModel is a visit history that includes only the node IDs of the visited locations visited in the previous step.

[0063] ≪Proposed technology when using method 2 for selecting the edge pair to be explained≫ The proposed technology (2-1) to (2-2), (2-3') to (2-5') can be realized as follows.

[0064] First, calculate the following:

[0065]

[0066] However, the following applies:

[0067]

[0068] In other words, the step number ^t and node ID ^c corresponding to the pair with the largest difference in recommendation degree among the pairs of real edges included in the real route and their corresponding counterreal virtual edges are calculated. Note that, in the above equation 12 as well, in order to ensure consistency between the VRP solver and the recommendation model RecModel, the input to the recommendation model RecModel is a visit history that includes only the node IDs of the visited locations visited in the previous step.

[0069] At this time, the actual edge e being explained fact The counterfactual edge e being explained cf Let the following be the conditions for each of them.

[0070]

[0071] The following describes an explanatory device 10 that, after selecting a pair of real edges and counterreal virtual edges to be explained using the proposed technology described above, can provide an explanation for those real edges and counterreal virtual edges (i.e., an explanation for questions such as "Why was this counterreal virtual edge selected instead of this real edge?").

[0072] <Example of Hardware Configuration of Explanation Device 10> Figure 3 is a diagram showing an example of the hardware configuration of an explanation device 10 according to one embodiment. As shown in Figure 3, the explanation device 10 according to one embodiment includes an input device 101, a display device 102, an external I / F 103, a communication I / F 104, a RAM (Random Access Memory) 105, a ROM (Read Only Memory) 106, an auxiliary storage device 107, and a processor 108. Each of these hardware components is connected to communicate via a bus 109.

[0073] The input device 101 is, for example, a keyboard, mouse, touch panel, physical buttons, etc. The display device 102 is, for example, a display, display panel, etc. Note that the explanatory device 10 does not necessarily have to have at least one of the input device 101 and the display device 102.

[0074] The external I / F 103 is an interface with external devices such as the recording medium 103a. Examples of recording media 103a include CDs (Compact Discs), DVDs (Digital Versatile Disks), SD memory cards (Secure Digital memory cards), and USB (Universal Serial Bus) memory cards.

[0075] The communication interface 104 is an interface for the demonstrator 10 to connect to a communication network. The RAM 105 is a volatile semiconductor memory (storage device) that temporarily holds programs and data. The ROM 106 is a non-volatile semiconductor memory (storage device) that can retain programs and data even when the power is turned off. The auxiliary storage device 107 is a non-volatile storage device such as an HDD (Hard Disk Drive), SSD (Solid State Drive), or flash memory. The processor 108 is a type of arithmetic unit such as a CPU (Central Processing Unit) or GPU (Graphic Processing Unit).

[0076] Note that the hardware configuration shown in Figure 3 is just one example, and the hardware configuration of the explanatory device 10 is not limited to this. The explanatory device 10 may, for example, have multiple auxiliary storage devices 107 or multiple processors 108, or it may not have some of the hardware shown, or it may have various hardware other than the hardware shown.

[0077] <Example of Functional Configuration of Explanation Device 10> Figure 4 is a diagram showing an example of the functional configuration of an explanation device 10 according to one embodiment. As shown in Figure 4, the explanation device 10 according to one embodiment has a route generation unit 201, an edge selection unit 202 for explanation, and an explanation unit 203. Each of these units is realized, for example, by a process in which one or more programs installed in the explanation device 10 are executed by a processor 108 or the like.

[0078] The route generation unit 201 takes a VRP instance defined as graph G = (V, E, X) as input and generates an actual route using a VRP solver or the like.

[0079] The edge selection unit 202 takes the actual route generated by the route generation unit 201 and the graph G = (V, E, X) as input and, using the proposed technology described above, selects the actual edge e to be explained. fact The counterfactual edge e being explained cf The selection unit 202 selects the actual edge e to be explained using the proposed technology which employs either method 1 or 2 for selecting the set of edges to be explained. fact The counterfactual edge e being explained cf You can choose either option.

[0080] The explanation unit 203 includes the actual route generated by the route generation unit 201, the graph G = (V, E, X), and the actual edge e selected by the edge selection unit 202. fact and counterfactual edge e cf Using and as input, the actual edge e described is explained by the prior art described in Non-Patent Document 1. fact and counterfactual edge e cf Generate and output a descriptive text for [the specified item].

[0081] In the example shown in Figure 1, a single explanation device 10 has a route generation unit 201, an edge selection unit 202 to be explained, and an explanation unit 203. However, these units may be distributed across a system composed of multiple devices. In this case, for example, the device on which the edge selection unit 202 to be explained is located may be called a "selection device" or the like.

[0082] <Example of explanatory process> Figure 5 is a flowchart showing an example of an explanatory process according to one embodiment.

[0083] The route generation unit 201 takes a VRP instance defined as graph G = (V, E, X) as input and generates an actual route using a VRP solver or the like (step S101).

[0084] The edge selection unit 202 takes the actual route generated in step S101 and the graph G input in step S101 as input and uses the proposed technology to select the actual edge e to be explained. fact The counterfactual edge e being explained cf Select (step S102). That is, when using the method 1 for selecting the edge set to be explained, the edge selection unit 202 calculates ^t using the above numbers 5 and 6, and ^c using the above numbers 8 and 9, thereby determining the e shown in the above number 7. fact The actual edge to be explained, e shown in the number 10 above. cf Select as the counterfactual virtual edge to be explained. On the other hand, when using the method 2 for selecting the set of edges to be explained, the edge selection unit 202 calculates ^t and ^c using equations 11 and 12, thereby obtaining the e shown in equation 13 above. fact and e cf Select the actual edge and counteracting virtual edge to be explained, respectively.

[0085] The explanation section 203 shows the actual route generated in step S101, the graph G input in step S101, and the actual edge e selected in step S102. fact and counterfactual edge e cf Using and as input, the actual edge e described is explained by the prior art described in Non-Patent Document 1. factand counterfactual edge e cf A descriptive text for this is generated and output (step S103).

[0086] <Modification> In the above embodiment, the actual route was generated from the VRP instance by a VRP solver or the like, but for example, the actual route may be generated from the VRP instance by a recommendation model RecModel. That is, the actual route may be generated by sequentially determining the destination locations from the VRP instance using a recommendation model RecModel.

[0087] <Summary> As described above, the explanation device 10 according to one embodiment calculates a recommendation score for each real edge included in the real route based on user attribute information representing the user's attributes and a recommendation model. Then, based on this recommendation score, the explanation device 10 according to one embodiment selects as the subject of explanation the pair of real edge with the lowest recommendation score and the counterreal virtual edge with the highest recommendation score, or the pair of real edge and counterreal virtual edge with the largest difference in recommendation scores. As a result, the explanation device 10 according to one embodiment can select the pair of real edge and counterreal virtual edge to be explained based, for example, on the user's subjective preferences.

[0088] Furthermore, prior art related to the technology described in Non-Patent Document 1 includes the technologies described in Reference Document 1 and Reference Document 2. By applying these technologies, it is possible to generate explanations for pairs of real edges and counterfactual virtual edges, but in all cases, the user must select the counterfactual virtual edges.

[0089] On the other hand, by applying the technology described in Reference 3, it is possible to select counterfactual virtual edges in the above modified example case, similar to the proposed technology when using Method 1 for selecting the edge set to be explained. However, if a real route is generated by a VRP solver or the like, it is not possible to select counterfactual virtual edges.

[0090] <Supplement> The functions of the elements disclosed herein may be implemented using circuits or processing circuits that include general-purpose processors, special-purpose processors, integrated circuits, ASICs (Application Specific Integrated Circuits), FPGAs (Field Programmable Gate Arrays), conventional circuits, and / or combinations thereof that are programmed using one or more programs stored in one or more memories, or otherwise configured to perform the disclosed functions. A processor is considered a circuit or processing circuit because it includes transistors and other circuits. A processor may be a programmed processor that executes programs stored in memory. In this disclosure, a circuit, unit, or means is hardware that performs the enumerated functions, or hardware programmed to perform the enumerated functions. Hardware may be any hardware disclosed herein that is programmed or configured to perform the enumerated functions.

[0091] There is a memory for storing a computer program that includes computer instructions. These computer instructions provide logic and routines that enable hardware (e.g., circuits or processing circuits) to perform the methods disclosed herein. This computer program can be implemented in commonly known forms as a computer-readable storage medium, a computer program product, a memory device, a recording medium such as a CD-ROM or DVD, and / or memory of an FPGA or ASIC.

[0092] The following additional information is disclosed regarding the embodiments described above.

[0093] (Note 1) A selection device for selecting a first edge included in a predetermined path on a graph and a second edge having the same starting point as the first edge but a different ending point, comprising: a memory and at least one processor connected to the memory, wherein the processor selects the first edge from among the edges constituting the path and the second edge from among the edges having the same starting point as the first edge but a different ending point, based on user attribute information representing the attributes of the user, the path, and information about each point included in the graph. (Note 2) The selection device according to Note 1, wherein the processor selects the first edge and the second edge using a model that takes the user attribute information, information about each point, and a point representing the starting point of the edges constituting the path as input and outputs a probability representing the recommendation level of the next point. (Note 3) A non-temporary storage medium storing a program that can perform a selection process by a computer, wherein the selection process is a selection process that selects a first edge included in a predetermined path on a graph and a second edge having the same starting point as the first edge but a different ending point, and the non-temporary storage medium selects the first edge from among the edges constituting the path and the second edge from among the edges having the same starting point as the first edge but a different ending point, based on user attribute information representing the user's attributes, the path, and information about each point included in the graph.

[0094] The present invention is not limited to the embodiments specifically disclosed above, and various modifications, changes, and combinations with known technologies are possible without departing from the spirit of the claims.

[0095] [References] Reference 1: van der Waa et al., Contrastive Explanations for Reinforcement Learning in terms of Expected Consequences, XAI workshop on the IJCAI 2018. Reference 2: Madumal et al., Explainable Reinforcement Learning through a Causal Lens, AAAI 2020. Reference 3: Amitai et al., Explaining RL Agents Through Counterfactual Action Outcomes, AAAI 2024.

[0096] 10 Explanation device 101 Input device 102 Display device 103 External I / F 103a Recording medium 104 Communication I / F 105 RAM 106 ROM 107 Auxiliary storage device 108 Processor 109 Bus 201 Route generation unit 202 Edge selection unit for explanation 203 Explanation unit

Claims

1. A selection device for selecting a first edge included in a predetermined path on a graph and a second edge having the same starting point as the first edge but a different ending point, the selection device comprising: a selection unit that selects the first edge from among the edges constituting the path and the second edge from among the edges having the same starting point as the first edge but a different ending point, based on user attribute information representing the user's attributes, the path, and information about each point included in the graph.

2. The selection device according to claim 1, wherein the selection unit selects the first edge and the second edge using a model that takes the user attribute information, the information about each point, and a point representing the starting point of the edge constituting the path as input and outputs a probability representing the recommendation level of the next point.

3. A selection method comprising a computer that selects a first edge included in a predetermined path on a graph and a second edge having the same starting point as the first edge but a different ending point, and which performs a selection procedure to select the first edge from among the edges constituting the path and the second edge from among the edges having the same starting point as the first edge but a different ending point, based on user attribute information representing the user's attributes, the path, and information about each point included in the graph.

4. A program that causes a computer to select a first edge included in a predetermined path on a graph and a second edge having the same starting point as the first edge but a different ending point, to execute a selection procedure that selects the first edge from among the edges constituting the path and the second edge from among the edges having the same starting point as the first edge but a different ending point, based on user attribute information representing the user's attributes, the path, and information about each point included in the graph.