Information processing device, information processing method, and program

Abstract

To provide an information processing apparatus, an information processing method, and a program that enable graph partitioning using a graph convolutional neural network without preparing teacher data.SOLUTION: An information processing apparatus 10 classifies a plurality of delivery destinations into a plurality of groups. The information processing apparatus includes: a learning section that performs unsupervised learning to train a graph convolutional neural network, which is determined using an adjacency matrix indicating a connection relationship of the plurality of delivery destinations and receives as input a feature matrix indicating a feature of the plurality of delivery destinations, using a first loss function defined such that the smaller a value for distance between delivery destinations belonging to a same group and the smaller a difference in features between delivery destinations belonging to a same group, the less a loss; and an output section that outputs information about a group to which the plurality of delivery destinations belongs, the information being obtained by inputting the feature matrix into the graph convolutional neural network trained by the learning section.SELECTED DRAWING: Figure 3

Description

[Technical field]

[0001] The present invention relates to an information processing device, an information processing method, and a program. [Background technology]

[0002] A graph partition is a set of vertices and edges connecting the vertices. The objective of this study is to divide a graph into multiple subsets. There are many known algorithms for graph partitioning. For example, the Kernighan-Lin algorithm, the Fiduccia-Mattheyses algorithm, Spectral Bisection method and others are also being used. Recently, neural networks have also been used. Algorithms using graph convolutional neural networks, which are a type of neural network, have also been proposed. (See, for example, Non-Patent Document 1). [Prior art documents] [Non-patent literature]

[0003] [Non-Patent Document 1] Thomas N. Kipf, Max Welling, Semi-Supervised Classification with Graph Convolutional Networks, [online], 2017, [Retrieved November 24, 2022], Internet<URL http: / / arxiv.org / abs / 1609.02907> Summary of the Invention [Problem to be solved by the invention]

[0004] When graph partitioning is performed using the technique described in Non-Patent Document 1, It is necessary to train the network using training data. Even if we try to use convolutional neural networks to solve real-world problems, In the first place, it may be difficult to prepare training data in advance.

[0005] Therefore, the present disclosure provides a method for constructing a graph convolutional neural network without preparing training data. Information processing device, information processing method, and program enabling graph partitioning using work The purpose is to provide. [Means for solving the problem]

[0006] An information processing device according to an aspect of the present disclosure is an information processing device for classifying a plurality of delivery destinations into a plurality of groups. A processing device, which is determined using an adjacency matrix indicating a connection relationship of the plurality of delivery destinations, A graph convolutional neural network that uses a feature matrix representing the characteristics of multiple delivery destinations as input. The smaller the value of the distance between delivery destinations belonging to the same group, the more likely it is that The first step is defined so that the loss is smaller as the difference in features between the delivery destinations in the group is smaller. A learning unit that performs unsupervised learning using a loss function, and Obtained by inputting the feature matrix into the graph convolutional neural network, and an output unit that outputs information regarding a group to which each of the plurality of delivery destinations belongs. Effect of the Invention

[0007] According to the present disclosure, a graph convolutional neural network can be constructed without preparing training data. We provide an information processing device, an information processing method, and a program that enable graph division using a graph It can be provided. [Brief description of the drawings]

[0008] [Figure 1] FIG. 1 is a diagram illustrating an example of a delivery management system according to an embodiment of the present invention. [Diagram 2] FIG. 2 is a diagram illustrating an example of a hardware configuration of an information processing device. [Diagram 3] FIG. 2 is a diagram illustrating an example of a functional block configuration of an information processing device. [Figure 4] 11 is a flowchart illustrating an example of a processing procedure executed by an information processing device. [Diagram 5] FIG. 13 is a diagram for explaining a graph. [Figure 6] FIG. 13 is a diagram showing the results of classifying a plurality of delivery destinations into a plurality of groups. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0009] An embodiment of the present invention will be described with reference to the accompanying drawings. Those marked with the same symbol have the same or similar configuration.

[0010] <System configuration> FIG 1 is a diagram showing an example of a delivery management system according to the present embodiment. The information processing device 10 and the terminal 20 are connected wirelessly or are connected via a wired communication network N and can communicate with each other.

[0011] The information processing device 10 is a device for managing the delivery of packages, and multiple delivery destinations are grouped into multiple groups. The information processing device 10 also performs a process of classifying (clustering) the data into groups. For example, the traveling salesman problem (traveling salesman p By using algorithms that solve the problem of stochastic randomness, it is possible to deliver packages efficiently. The information processing device 10 may determine a delivery route. It may be composed of a virtual server or a virtual machine that runs on a hypervisor. It may be configured using a general-purpose server or a cloud server. good.

[0012] The terminal 20 is a terminal operated by a user, such as a smartphone or a tablet terminal. , a mobile phone, a personal computer (PC), a notebook PC, etc. Various data output from the information processing device 10 is displayed on the screen. The information processing device 10 can be operated via the terminal 20.

[0013] The information processing device 10 divides a plurality of delivery destinations into a plurality of groups using a graph partitioning algorithm. Specifically, the information processing device 10 classifies the graph convolution neural network into By using a Graph Convolutional Neural Network (GCN), the system classifies delivery destinations. In the following, we will refer to the graph convolutional neural network simply as the graph convolutional network. The network is called GCN.

[0014] Conventionally, when using GCN to perform graph partitioning, training data is prepared in advance and the GCN is trained. However, since the delivery destination of the parcels changes every day, it was necessary to train the teacher in advance. Therefore, in this embodiment, the information processing device 10 uses the GC By devising a loss function to be used when training N, We trained the GCN without using

[0015] <Hardware configuration> FIG. 2 is a diagram illustrating an example of a hardware configuration of the information processing device 10. CPU (Central Processing Unit), GPU (Graphical Processing Unit) and other processors processor 11, memory (e.g., RAM or ROM), HDD (Hard Disk Drive) and / or The storage device 12 such as an SSD (Solid State Drive), a network for wired or wireless communication, A network interface (IF) 13, an input device 14 for receiving input operations, and an information output device The input device 14 includes, for example, a keyboard, a touch panel, The output device 15 is, for example, a display, a touch panel, etc. and / or speakers, etc.

[0016] <Function block configuration> FIG. 3 is a diagram showing an example of a functional block configuration of the information processing device 10. The storage unit 100 includes a storage unit 100, a receiving unit 101, a learning unit 102, and an output unit 103. 00 can be realized using the storage device 12 included in the information processing device 10. The reception unit 101, the learning unit 102, and the output unit 103 are integrated into the processor 1 of the information processing device 10. 1 can be realized by executing a program stored in the storage device 12. The program can be stored in a storage medium. The storage medium is a non-transitory computer-readable storage medium. The non-transitory storage medium may be, but is not limited to, a non-transitory readable medium. For example, a USB (Universal Serial Bus) memory or a CD-ROM (Compact Disc Read-On The storage medium may be a storage medium such as a portable memory.

[0017] The storage unit 100 stores various data related to delivery destinations (hereinafter referred to as "delivery destination data"), The learning model is stored in the GCN. The learning model contains information that determines the model structure of the GCN and various The parameter values ​​are included.

[0018] The reception unit 101 receives various data inputs from the terminal 20. For example, 1 accepts input of delivery destination data, etc.

[0019] The learning unit 102 trains a learning model using the delivery destination data and a predetermined loss function. Specifically, the learning unit 102 uses an adjacency matrix that indicates the connection relationship between multiple destinations. The GCN is input with a feature matrix that indicates the characteristics of the multiple delivery destinations, and the GCN is then used to identify the delivery destinations belonging to the same group. The smaller the value of the distance between the destinations in the same group, the greater the probability of The loss function (hereafter referred to as the "first loss") is defined so that the smaller the difference in features, the smaller the loss. The learning is performed using unsupervised learning, using a set of functions called "parameters".

[0020] The output unit 103 outputs information output from the learning model. 3 is a plurality of GCNs obtained by inputting a feature matrix to the GCN trained by the training unit 102. Prints information about the groups to which each of the destinations belongs.

[0021] <Processing Procedure> FIG. 4 is a flowchart illustrating an example of a processing procedure executed by the information processing device 10. 4, the information processing device 10 trains the GCN without using teacher data, A method for classifying destinations will now be described in detail.

[0022] In step S10, the receiving unit 101 receives input of delivery destination data. For example, the location information of the delivery destination (latitude and longitude, etc.), the desired delivery time period (for example, 14:00 to 16:00, etc.), and information on the relationship between the delivery destination and the delivery base (Depot), etc. The unit 101 stores the received delivery destination data in the storage unit 100 .

[0023] In step S11, the learning unit 102 derives an adjacency matrix and and generate a feature matrix.

[0024] Here, we will explain about graphs. A graph is a collection of multiple vertices and edges connecting the vertices. The graph is composed of a set of nodes, and can be expressed by the formula 1. Note that the graph is also called a graph network. Also called.

number

[0025] An example of a graph is shown in FIG. 5. The graph shown in FIG. 5 has five vertices (v1 to v5) and Seven edges connecting the points (e12, e14, e23, e24, e34, e35, e45,) In addition, a feature value can be associated with each vertex. In this mode, the features of the vertices, edges, and vertices are respectively the destinations, the distance between destinations, and the destinations. We define a graph based on the characteristics of each item and use GCN to partition the graph. Divide into groups of several.

[0026] Next, the adjacency matrix A will be explained. The adjacency matrix A is It indicates the connection relationship. When it is 1, the edges are connected, and when it is 0, the edges are not connected. In this embodiment, the element A of the adjacency matrix A is ij is expressed as Equation 2. In addition, i and j indicate the delivery destination. For example, if there are 100 delivery destinations, In this case, i and j are each represented by an integer of 1 to 100.

number

[0027] The value of θ is arbitrary. For example, if the delivery destinations are far apart and the same group If the distance between destinations i and j is θ If it is greater than the value of A ij Since ,is 0, there is a will be represented as having no edges connected.

[0028] In addition, even if the distance between two delivery destinations is less than or equal to θ, When there is an obstacle (such as a river) that delivery vehicles (trucks, etc.) cannot pass through In this case, the learning unit 102 selects A ij Make sure to set the value to 0. For example, the reception unit 101 may receive a notification from the user that a fault exists between two delivery destinations. Upon receiving the designation, the learning unit 102 selects the two elements of the adjacency matrix that are received by the receiving unit 101. A corresponding between two delivery destinations ij The value may be set to 0.

[0029] Next, we will explain the degree matrix D. The degree matrix D is the number of vertices that go to each vertex. The diagonal matrix D indicates which edges are connected. Component D ij can be expressed by Equation 3.

number

[0030] Next, we will explain the feature matrix X. The feature matrix X is For example, the number of vertices is 1 to n, and i is the identifier of the feature. Then, the feature quantities X of each of vertices 1 to n are i is expressed by the following Equation 4.

number

number

[0031] In addition, the feature associated with the delivery destination is the operation time of the delivery vehicle during the desired delivery time period. information on the proportion of time that includes working hours (which may be called the "overlap rate"); In other words, the feature matrix X may include the desired delivery time as the feature of multiple delivery destinations. Information regarding the percentage of the time slot during which the delivery vehicle is in operation may also be included. This information is expressed as a value between 0 and 1. If it is 0, the delivery vehicle will not be available during the desired delivery time. In addition, if the value is 1, all the delivery times are included. This indicates that the time period is included in the operating hours of the delivery vehicle. For example, if the desired delivery time is 8:00 If the delivery time is from 10:00 to 12:00 and the delivery vehicle is operating from 10:00 to 16:00, the desired delivery time ( Of the 4 hours, the time during which the delivery vehicle is in operation is 2 hours (10:00-12:00) So, 2 hours ÷ 4 hours = 0.5.

[0032] In addition, the feature associated with the delivery destination includes the direction from the delivery depot to the delivery destination ( or vice versa) and information regarding the distance between each of the multiple delivery destinations and the delivery center In other words, the feature matrix X may include the following features of multiple delivery destinations: Information regarding the direction to the delivery destination or the direction from the delivery destination to the delivery base, and each of multiple delivery destinations The distance information may include information about the distance between the delivery point and the delivery location. is the Euclidean distance, the havesine distance or the It may be the actual distance on the diagram.

[0033] The feature quantities included in the feature matrix X are not limited to those mentioned above. Any feature quantity related to the delivery destination can be included. Such features may be included.

[0034] Here, specific examples of the adjacency matrix A, the degree matrix D, and the feature matrix X are shown in Equations 6 to 8. The matrices shown in Equations 6 to 8 are graphs with three vertices and edges connecting each vertex. handle.

number

number

number

[0035] In step S12, the learning unit 102 performs learning of the GCN using a loss function. For example, the learning unit 102 may perform live learning for training a neural network. By setting the GCN model and the loss function to be used in the library, the GCN can be trained. Equation 9 shows an example of the GCN according to this embodiment.

number

number

[0036] In this embodiment, learning is performed using a GCN with two hidden layers. where Equation 12 represents the first hidden layer, and Equation 13 represents the output layer.

number

number

number

[0037] As shown in Equation 11, the elements of the feature matrix are input to the input layer of the GCN. The output layer of N outputs the probability that each delivery destination belongs to each group. For example, If the number of is 100 and the number of possible groups is 5, then there are 5 probabilities for each destination. (Probability of belonging to group 1, probability of belonging to group 2, probability of belonging to group 3, The probability of belonging to group 4, the probability of belonging to group 5, etc. are output. The large value may be 1, but the present embodiment is not limited to this.

[0038] Next, the learning unit 102 performs a process of learning the GCN using the loss function. As shown above, the learning unit 102 determines whether the distance between delivery destinations belonging to the same group is smaller. The smaller the difference in features between delivery destinations in the same group, the smaller the loss. We perform unsupervised learning using the first loss function defined as follows. By using this method, the distance between destinations is short and destinations with similar characteristics are grouped together. The GCN can be trained to belong to a group. Equation 14 and Equation 15 show an example of the first loss function.

number

number

[0039] The first half of Equation 15, “λ1γ ij The smaller the distance between destination i and destination j, the The latter part of Equation 15 (λ1γ ij The rest of the part is the shipping destination i and The smaller the difference between each feature value between destinations i and j, the smaller the value. If destination j belongs to the same group, P ik ×P jk The value of becomes large. Therefore, P in Equation 14 ik ×P jk ×ω ij The value of is the number of destinations i and The smaller the distance between the destinations, the smaller the value. It can be seen that the smaller the difference in the feature between j, the smaller the value becomes.

[0040] In addition to the first loss function, the learning unit 102 also calculates the probability that each delivery destination belongs to a predetermined group. The value based on the difference between the total value of and the average number of delivery destinations per group is calculated for multiple groups. The second loss is defined so that the smaller the total value of each calculated value, the smaller the loss. A loss function may be used to perform unsupervised learning. Based on the difference between the total probability of belonging to a given group and the average number of delivery destinations per group The "value" may be the square of the difference, the absolute value of the difference, or the The square root of the squared loss function can be used. We can train the GCN so that the number of is averaged across groups. Equation 16 can be written as An example of the second loss function is shown below.

number

[0041] In addition to the first loss function, the learning unit 102 also calculates whether each delivery destination is one of multiple groups. The maximum probability of belonging to is the maximum possible value for probability (e.g., 1, but not limited to this). The loss function is defined as the loss is smaller the closer the training set is to the training set (not specified). By using the third loss function, the delivery destination can be GC is a method to determine the probability of belonging to a group so that the maximum probability approaches the maximum possible value. For example, before training the GCN, a certain delivery destination is Suppose the output shows that the probabilities of belonging to groups 1 to 3 are (0.3, 0.3, 0.4). In this case, by training the GCN using the third loss function, for example (0.1, 0.1, 0.8), it is possible to obtain output results that clearly show which group the item belongs to. An example of the third loss function is shown in Equation 17.

number

[0042] In addition to the first loss function, the learning unit 102 uses a second loss function and a third loss function. In the above, learning may be performed by unsupervised learning. An example of a loss function when the second loss function and the third loss function are used is shown below.

number

[0043] <Example> FIG. 6 shows the result of classifying multiple delivery destinations into multiple groups. In the example of Figure 6, multiple delivery destinations are classified into 10 groups. It is clear that there are

[0044] <Summary> According to the embodiment described above, by training a GCN using only a loss function, It is now possible to classify multiple delivery destinations into multiple groups without the need for training data. become.

[0045] The above-described embodiment is intended to facilitate understanding of the present invention and does not limit the present invention. The flowcharts, sequences, and implementations described in the embodiments are not intended to be interpreted as being related to the present invention. The elements of the embodiment, as well as their arrangement, materials, conditions, shapes, sizes, etc. are merely examples. The present invention is not limited to the above and may be modified as appropriate. The components may be partially substituted or combined.

[0046] <Additional Notes> This embodiment may be expressed as follows.

[0047] <Appendix 1> An information processing device that classifies a plurality of delivery destinations into a plurality of groups, The characteristics of the plurality of delivery destinations are determined using an adjacency matrix that indicates a connection relationship between the plurality of delivery destinations, A graph convolutional neural network is applied to the same group using a feature matrix that indicates the characteristics of the The smaller the value of the distance between delivery destinations in the same group, the more likely it is that delivery destinations in the same group will be distributed. The first loss function is defined so that the smaller the difference in features between the previous images, the smaller the loss. A learning unit that performs learning by unsupervised learning; The feature row is added to the graph convolutional neural network trained by the training unit. Information regarding the group to which each of the multiple delivery destinations belongs, obtained by inputting a column an output unit for outputting the signal; An information processing device having the above configuration.

[0048] <Appendix 2> The learning unit calculates, in addition to the first loss function, a sum of probabilities that each delivery destination belongs to a predetermined group. A value based on the difference between the total value and the average number of delivery destinations per group is calculated for the plurality of groups. The second loss is defined so that the smaller the total value of each calculated value, the smaller the loss. The unsupervised learning is performed using a loss function. 2. An information processing device according to claim 1.

[0049] <Appendix 3> The learning unit is configured to, in addition to the first loss function, determine whether each delivery destination is one of the plurality of groups. The closer the maximum probability of belonging to is to the maximum possible value of the probability, the smaller the loss. The unsupervised learning is performed using a third loss function defined as follows: 2. An information processing device according to claim 1.

[0050] <Appendix 4> In addition to the first loss function, the learning unit The sum of the probability that each destination belongs to a given group and the average number of destinations per group The value based on the difference between the A second loss function is defined so that the loss is smaller as The maximum probability that each delivery destination belongs to one of the multiple groups is set as the probability. The third loss function is defined so that the closer the value is to the maximum possible value, the smaller the loss becomes. The unsupervised learning is performed using 2. An information processing device according to claim 1.

[0051] <Appendix 5> The feature matrix includes information related to desired delivery time periods as features of the plurality of delivery destinations. Information included, 5. The information processing device according to any one of claims 1 to 4.

[0052] <Appendix 6> The feature matrix includes, as features of the plurality of delivery destinations, desired delivery time zones and delivery vehicles. information about the percentage of overlap with the operating hours of 6. The information processing device according to any one of claims 1 to 5.

[0053] <Appendix 7> The feature matrix may include, as features of the plurality of delivery destinations, a direction from a delivery base to a delivery destination or or information on a direction from a delivery destination to the delivery base, and and information regarding the distance between the delivery point and the 7. The information processing device according to claim 1,

[0054] <Appendix 8> An information processing method executed by an information processing device for classifying a plurality of delivery destinations into a plurality of groups. hand, The characteristics of the plurality of delivery destinations are determined using an adjacency matrix that indicates a connection relationship between the plurality of delivery destinations, A graph convolutional neural network is applied to the same group using a feature matrix that indicates the characteristics of the The smaller the value of the distance between delivery destinations in the same group, the more likely it is that delivery destinations in the same group will be distributed. The first loss function is defined so that the smaller the difference in features between the previous images, the smaller the loss. A step of performing learning by unsupervised learning; inputting the feature matrix into the trained graph convolutional neural network; outputting information on a group to which each of the plurality of delivery destinations belongs, the information being obtained in the step and, An information processing method comprising:

[0055] <Appendix 9> A computer that classifies multiple delivery destinations into multiple groups. The characteristics of the plurality of delivery destinations are determined using an adjacency matrix that indicates a connection relationship between the plurality of delivery destinations, A graph convolutional neural network is applied to the same group using a feature matrix that indicates the characteristics of the The smaller the value of the distance between delivery destinations in the same group, the more likely it is that delivery destinations in the same group will be distributed. The first loss function is defined so that the smaller the difference in features between the previous images, the smaller the loss. A step of performing learning by unsupervised learning; inputting the feature matrix into the trained graph convolutional neural network; outputting information on a group to which each of the plurality of delivery destinations belongs, the information being obtained in the step and, A program for executing. [Explanation of symbols]

[0056] 1...Delivery management system, 10...Information processing device, 11...Processor, 12...Storage device, 1 3... network IF, 14... input device, 15... output device, 20... terminal, 100... storage unit 101: Reception unit; 102: Learning unit; 103: Output unit

Claims

1. An information processing device that classifies a plurality of delivery destinations into a plurality of groups, a learning unit that performs unsupervised learning of a graph convolutional neural network that is determined using the connection relationships of the plurality of delivery destinations and that receives information indicating characteristics of the plurality of delivery destinations as an input, using a loss function that is defined so that the smaller the value related to the distance between delivery destinations that belong to the same group and the smaller the difference in characteristics between delivery destinations that belong to the same group, the smaller the loss; an output unit that outputs information regarding a group to which each of the plurality of delivery destinations belongs, the information being obtained by inputting information indicating characteristics of the plurality of delivery destinations to the graph convolutional neural network trained by the learning unit; An information processing device having the above configuration.

2. The information indicating the characteristics of the plurality of delivery destinations includes information regarding a desired time period for delivery as a characteristic of the plurality of delivery destinations. The information processing device according to claim 1 .

3. The information indicating the characteristics of the plurality of delivery destinations includes, as the characteristics of the plurality of delivery destinations, information regarding the percentage of overlap between the desired delivery time period and the operating hours of the delivery vehicle. The information processing device according to claim 2 .

4. The information indicating the characteristics of the plurality of delivery destinations includes, as the characteristics of the plurality of delivery destinations, information regarding a direction from a delivery base to the delivery destination or a direction from the delivery destination to the delivery base, and information regarding a distance between each of the plurality of delivery destinations and the delivery base. The information processing device according to claim 1 .

5. An information processing method executed by an information processing device for classifying a plurality of delivery destinations into a plurality of groups, comprising: A step of performing unsupervised learning on a graph convolutional neural network that is determined using the connection relationships of the plurality of delivery destinations and that inputs information indicating characteristics of the plurality of delivery destinations, using a loss function that is defined so that the loss decreases as the value related to the distance between delivery destinations belonging to the same group decreases and as the difference in characteristics between delivery destinations belonging to the same group decreases; A step of outputting information regarding a group to which each of the plurality of delivery destinations belongs, the information being obtained by inputting information indicating characteristics of the plurality of delivery destinations to the trained graph convolutional neural network; An information processing method comprising:

6. A computer that classifies multiple delivery destinations into multiple groups. A step of performing unsupervised learning on a graph convolutional neural network that is determined using the connection relationships of the plurality of delivery destinations and that inputs information indicating characteristics of the plurality of delivery destinations, using a loss function that is defined so that the loss decreases as the value related to the distance between delivery destinations belonging to the same group decreases and as the difference in characteristics between delivery destinations belonging to the same group decreases; A step of outputting information regarding a group to which each of the plurality of delivery destinations belongs, the information being obtained by inputting information indicating characteristics of the plurality of delivery destinations to the trained graph convolutional neural network; A program for executing the above.

Images