JP7844262B2 - Recommendation generation system, recommendation generation method, and recommendation generation program - Google Patents

Description

This invention relates to a recommendation generation system, a recommendation generation method, and a recommendation generation program.

As background technology for this field, there is Japanese Patent Publication No. 2021-135722 (Patent Document 1). This publication states, "We provide an information processing device and program that can quantify and make available for use the relationships between information items such as the subject of analysis and products, even in social media where the relationships are loose."

Japanese Patent Publication No. 2021-135722

In recent years, technologies that recommend urban functions suitable for specific locations such as various facilities and urban spaces to businesses and local governments that manage and provide urban functions, and that support the introduction or repurposing of urban functions, have been attracting attention.

One possible method for realizing this technology is to obtain location information of urban functions from map data, analyze the locational relationships between urban functions, and generate recommendations based on the high affinity between individual locations and surrounding urban functions.

The technology described in Patent Document 1 constructs a graph in which social media users are represented as nodes and direct connection relationships between users on social media are represented as edges. This allows for the quantification of relationships between all nodes, including those without direct connections, using machine learning techniques. By applying the technology described in Patent Document 1, it is also possible to construct a graph in which urban functions implemented at a given location are represented as nodes and urban functions located within a predetermined distance are represented as edges. By applying a predetermined machine learning technique, the relationships between all urban functions can be indirectly quantified. Furthermore, the affinity between candidate urban functions for recommendation and surrounding urban functions can be calculated from the estimated relationships between all urban functions and a list of urban functions located within a predetermined distance from the recommendation target location. This allows for the quantification of the affinity between all candidate urban functions and surrounding urban functions. By sorting urban functions in order of affinity, it is possible to generate recommendations for urban functions at a given location.

However, in this recommendation generation method, it is necessary to pre-set a predetermined distance for the range of surrounding urban functions to be considered when calculating affinity with the recommended locations. Generally, system users do not have information about the range of surrounding urban functions that could potentially create affinity with the recommended locations, making it difficult to set an appropriate distance.

Furthermore, in this recommendation generation method, the graph used as input for the machine learning method is constructed using only the current spatial location relationships of urban functions. Therefore, the relationships between urban functions estimated by the machine learning method only indicate the strength of the locational correlation between urban functions, and do not consider how urban function users evaluate the relationships between urban functions. As a result, this recommendation generation method may recommend urban functions that have high locational affinity with surrounding urban functions but are undesirable to urban function users. To prevent this, it is necessary to consider value information regarding the location of urban functions.

The present invention aims to provide a technology that enables recommendations that take into account value information regarding the location of urban functions.

To solve the above problems, for example, the configuration described in the claims may be adopted.
The present invention includes multiple means for solving the above problems, but to give one example, a recommendation generation system that uses a computer having a processor and memory to present urban functions for a location, wherein the processor is configured as a recommendation generation system characterized by: setting up nodes in a graph that points to the first location information and the second location information using a first table that associates first location information and value information, and a second table that associates second location information containing at least one of a plurality of first location information with urban functions, and constructing a single graph in which the urban functions are represented as nodes that connect to the first location information and the value information as nodes that connect to the second location information; a graph learning process that expresses the interrelationships between nodes in a vector space using a machine learning method based on the graph; and a recommendation result generation process that evaluates the path length from the node distributed representation vectors of the first location information, the urban functions, and the value information based on the specification of the first location information, evaluates the path length between nodes from the vector space, and generates a recommendation for the urban functions for the first location information.

According to one aspect of the present invention, it is possible to provide a technology that enables recommendations that take into account value information regarding the location of urban functions.

This is a system configuration diagram of the recommendation generation system in the example. This is a hardware configuration diagram for realizing the recommendation generation system of the embodiment. This figure shows an example of a value information table. This figure shows an example of a table of implemented urban functions. This figure shows an example of a table representing a spatial structure. This figure shows an example of a graph-structured table (node table). This figure shows an example of a graph-structured table (edge table). This figure shows an example of a distributed representation table. This figure shows an example of a recommendation results table. This figure shows an example of a screen used to perform the initial setup for graph construction. This figure shows an example of a screen for setting the locations to be recommended. This figure shows an example of a screen for displaying recommendation results. This is a flowchart illustrating the operation for generating recommendations in Example 1. This figure shows an example of the screen used to perform the initial setup for graph construction in Practical Example 2. This is a flowchart illustrating the operation for generating recommendations in Example 2. This figure shows an example of a screen used to perform the initial setup for graph construction in Example 3. This is a flowchart illustrating the operation for generating recommendations in Example 3.

The following embodiments will be described with reference to the drawings. Note that the embodiments described below are not intended to limit the invention as defined in the claims, and not all of the elements and combinations thereof described in the embodiments are necessarily essential to the solution of the invention.

In other words, the following description and drawings are illustrative examples for explaining the present invention, and have been omitted and simplified as appropriate for clarity of explanation. The present invention can also be implemented in various other forms. Unless otherwise specified, each component may be singular or plural.

The positions, sizes, shapes, and ranges of the components shown in the drawings may not represent their actual positions, sizes, shapes, and ranges in order to facilitate understanding of the invention. Therefore, the present invention is not necessarily limited to the positions, sizes, shapes, and ranges disclosed in the drawings.

In the following explanation, various types of information may be described using terms such as "database," "table," and "list," but these types of information may also be represented using other data structures. To indicate independence from data structure, "XX table," "XX list," etc., may sometimes be referred to as "XX information." When describing identification information, if terms such as "identification information," "identifier," "name," "ID," and "number" are used, these terms are interchangeable.

When there are multiple components with the same or similar function, they may be described using the same symbol but with different subscripts. However, if it is not necessary to distinguish between these multiple components, the subscripts may be omitted in the description.

Furthermore, while the following explanation may describe the processing performed by executing a program, the processor (e.g., CPU (Central Processing Unit), GPU (Graphics Processing Unit)) executes the defined processing, using memory resources (e.g., memory) and/or interface devices (e.g., communication ports) as appropriate. Therefore, the processor may be the primary driver of the processing. Similarly, the primary driver of the processing performed by executing a program may be a controller, device, system, computer, or node having a processor. The primary driver of the processing performed by executing a program may be an arithmetic unit or arithmetic device, and may include dedicated circuits for specific processing (e.g., FPGA (Field-Programmable Gate Array) or ASIC (Application Specific Integrated Circuit)).

The program may be installed from the program source code into a device such as a computer. The program source code may be, for example, a program distribution server or a computer-readable storage medium. If the program source code is a program distribution server, the program distribution server includes a processor and storage resources for storing the program to be distributed, and the processor of the program distribution server may distribute the program to other computers. Furthermore, in the following description, two or more programs may be implemented as a single program, or one program may be implemented as two or more programs.

This embodiment describes an example of a recommendation generation method in the recommendation generation system A1100, which simultaneously evaluates the affinity with surrounding urban functions and the relationship with value indicators for a given location, and then recommends urban functions.

The recommendation generation system of this embodiment will be described with reference to Figure 1.
The recommendation generation system A1100 includes a value information DB (Database) A1111, an implemented urban function DBA1112, a spatial structure DBA1113, a graph structure DBA1114, a distributed representation DBA1115, a recommendation result DBA1116, an automatic data update unit A1121, a graph construction unit A1122, a graph learning unit A1123, a recommendation result generation unit A1124, and a user interface A1125.

Here, the Value Information DBA1111 is a database for storing and linking value information characteristic of each area, obtained through various statistical data and surveys conducted on residents, with the aggregated area. To enable linking with other databases within the Recommendation Generation System A1100, each record is assigned an area ID that uniquely identifies the area name.

The Implementing Urban Functions DBA1112 is a database published by mapmakers and public institutions that publish map information, containing coordinates and urban functions implemented at those coordinate points. To enable linking with other databases within the recommendation generation system A1100, each record is assigned a unique location ID that identifies the coordinate point.

The spatial structure DBA1113 is information used within the recommendation generation system A1100 to link the area IDs included in the value information DBA1111 and the location IDs included in the implemented urban function DBA1112, based on the hierarchy of urban space. It also holds the area names corresponding to the area IDs.

The graph structure DBA1114 is information used to record the structure of a single graph containing various information items necessary for generating recommendation results, such as location, area, urban function, and value information. It consists of a node table and an edge table.

The distributed representation DBA1115 is information used to record the characteristics of each graph node, obtained by inputting a graph into a machine learning method. It holds the node ID and a distributed representation of the graph node's characteristics as a high-dimensional real-valued vector.

The recommendation result DBA1116 is information used to record the results of evaluating the relationships between locations, urban functions, and value information from the node-distributed representation, using path length. In addition to urban functions and their path lengths, it also stores various information used to calculate the path length.

The automatic data update unit A1121 is a function for updating the information used in this system. It collects data via the local network or the internet (Network A1300) and registers the information in the value information DBA1111, the implemented urban function DBA1112, and the spatial structure DBA1113.

The graph construction unit A1122 is a function that enables the quantification of relationships using machine learning methods by constructing a single graph that represents the connection relationships of locations, areas, urban functions, and value information necessary for recommendation generation. It performs graph construction processing using value information DBA1111, implemented urban function DBA1112, and spatial structure DBA1113, and outputs the results to graph structure DBA1114. Initial settings for graph construction are performed by the system user using a screen as shown in Figures 9, 13, or 15.

The graph learning unit A1123 is a function for obtaining distributed representations for all nodes from the graph constructed by the graph construction unit A1122. It takes the graph structure DBA1114 as input, applies machine learning techniques to learn the distributed representations, and outputs the results to the distributed representation DBA1115.

The recommendation result generation unit A1124 takes the distributed representation DB1115 as input and is a function for evaluating the relationship between a location, urban functions, and value information using path length in a vector space. It outputs the evaluation result of urban functions for a location, considering its affinity with surrounding urban functions and its relationship with value information, to the recommendation result DBA1116.

When the user interface A1125 receives information from the user terminal A1200, it transmits that data to the graph construction unit A1122, graph learning unit A1123, and recommendation result generation unit A1124 of the recommendation generation system A1100.

The user terminal A1200 is a terminal used by the user to configure settings for the recommendation generation system A1100, and includes devices such as PCs and mobile phones. The user connects to the recommendation generation system A1100 using this terminal.

Referring to Figure 2, the hardware configuration for realizing the embodiment will be described.
The hardware configuration shown in Figure 2 is the hardware configuration H2000 for implementing the recommendation generation system A1100 and the user terminal A1200.

The hardware consists of a central processing unit (CCU) H2002, main memory (PMS) H2003, internal bus H2004, bus interface H2005, external bus H2006, screen output interface H2007, user input interface H2008, mass storage H2009, communication interface H2010, and input/output interface H2011.

The central processing unit H2002 is a device for performing calculations such as program execution, and is used, for example, to realize the operation of the graph construction unit A1122, graph learning unit A1123, and recommendation result generation unit A1124 of the recommendation generation system A1100 shown in Figure 1.

The main memory H2003 is used as a processing area during program execution and as a temporary data storage area. For example, it temporarily stores basic programs such as the OS (Operating System), as well as programs and temporary information necessary for processing on each device, such as the recommendation generation system A1100 and the user terminal A1200.

The central processing unit H2002 and the main memory H2003 are connected by an internal bus H2004, which in turn is connected to the external bus H2006 via a bus interface H2005.

The external bus H2006 is connected to the screen output device I/F H2007, user input device I/F H2008, mass storage device I/F H2009, communication device I/F H2010, and input/output device I/F H2011. Data input/output between these interfaces and the central processing unit H2002 and main memory H2003 is mediated using the bus I/F H2005 and internal bus H2004.

The screen output interface H2007 is an interface for connecting to the display H2100, and outputs information obtained via the external bus H2006 to the display H2100. The display H2100 is a device capable of displaying, for example, the settings screen shown in Figure 9, and can also display information for operating the basic program.

The User Input Device I/F H2008 is an interface for connecting to the User Input Device H2200 and outputs information received from the User Input Device to the external bus H2006. The User Input Device is a device that receives input from the user, such as a keyboard or mouse, and the information received from the User Input Device is output to the User Input Device I/F H2008.

The Mass Storage Device I/F H2009 is an interface for connecting to the Mass Storage Device H2300, and mediates data input/output from the Mass Storage Device to the external bus H2006. The Mass Storage Device is a device such as an HDD (Hard Disk Drive), and stores the basic program for realizing the functions of the Recommendation Generation System A1100, temporary storage information of program execution processing results, programs for realizing processing on each device such as the Recommendation Generation System A1100 and the user terminal A1200, and information such as value information DBA1111, implemented city function DBA1112, spatial structure DBA1113, graph structure DBA1114, distributed representation DBA1115, and recommendation result DBA1116 in the Recommendation Generation System A1100 when the device is powered off or when the program is not running. Furthermore, during the execution of various processes, the central processing unit H2002 reads these programs and data into the main memory H2003 and executes the programs. The mass storage device may be a recording device such as an SSD (Solid State Drive), or a recording medium such as an IC card, SD card, or DVD.

The communication device interface H2010 is an interface for connecting to the communication device H2400, and mediates data input/output from the communication device to the external bus H2006. The communication device is for connecting to external server devices, for example, via Ethernet®, and can communicate with systems connected via the network A1300 between recommendation generation systems A1100.

The I/O device interface H2011 is an interface for connecting to the I/O device H2500, and mediates data input/output from the I/O device to the external bus H2006. The I/O device includes devices such as drive devices capable of reading and writing to external media and finger vein readers for user authentication.

Figure 3 shows an example of a table for value information DBA1111. It includes the value information name T10, the area name T11 associated with the value information, a score T12 indicating the relative evaluation value of the value information for that area name, an area ID T13 to uniquely identify the area name, and the prefecture T14 to which the area belongs. By setting these values, it is possible to associate area names with value information. For example, "Area Name: XX Town" corresponds to the value index "Value Information: Child-rearing Ease." In Figure 3, for example, "XX Town," identified by area ID "1" belonging to the prefecture "Kanagawa Prefecture," shows that the score value for "Child-rearing Ease" is "10." This score value could be set, for example, to a percentage representing the child-rearing ease obtained from a survey.

Figure 4 shows an example of the table for the implemented urban function DBA1112. It includes a location ID T20 to uniquely identify a location, coordinates T21 to specify the location of the location, the implemented urban function T22 of the location, and the prefecture T23 to which the location belongs. By setting these values, it is possible to associate a location with an implemented urban function, for example, "Location ID: 1" with "Implemented Urban Function: Convenience Store". For example, in Figure 4, the coordinates of the location identified by location ID "1" belonging to the prefecture "Kanagawa Prefecture" are "(36.30, 140.46)", indicating the presence of a "convenience store".

Figure 5 shows an example of a table for the spatial structure DBA1113. It includes the location IDT31 defined in Figure 4, the area IDT32 defined in Figure 3, and the area name T33. By setting these values, it is possible to search for location IDs belonging to an area ID. In Figure 5, for example, it shows that in the prefecture "Kanagawa Prefecture," the location identified by location ID "1" and the location identified by location ID "2" belong to "XX Town," which is identified by area ID "1."

Figures 6A and 6B show examples of tables for the graph structure DBA1114. The graph structure DBA1114 manages the nodes and edges, which are components of the graph, using a node table and an edge table, respectively.

The node table shown in Figure 6A is used to record a list of nodes that make up the graph, and includes the node ID T41, node name T42, and node attribute T43. The node ID uniquely identifies a node, and multiple nodes will never have the same node ID. The node name describes the information item represented by the node, and multiple nodes may have the same node name. The node attribute describes the attribute of the information item represented by the node name, and can take one of the following values: "Value Information," "Urban Function," "Location," or "Area." In Figure 6A, for example, the node identified by node ID "1" with the node name "Livability" is shown to belong to the node attribute "Value Information."

The edge table shown in Figure 6B is used to record the connection relationships between nodes and includes edge ID 51 and edge constituent nodes T52. Furthermore, if information regarding the strength of the edge connection is available, this information can be represented by setting a numerical value for the edge weight T53 in the edge table. In Figure 6B, for example, the edge identified by edge ID "2" records the connection relationship between the node identified by node ID "3" and the node identified by node ID "4," and no edge weight is set.

Figure 7 shows an example of a distributed representation DBA1115 table. It includes node ID T61 as defined in Figure 6A, and distributed representation T62, which represents the characteristics of the node ID as a high-dimensional real-valued vector. Figure 7 shows, for example, that a node identified by node ID "1" is represented by the distributed representation (20.73, 28.56, 20.85, ...).

Figure 8 shows an example of a recommendation result DBA1116 table. It holds the urban function T71, which is a recommendation candidate, the route length T72, which is the evaluation result of that urban function, and the information T73 and T74 used to calculate the route length. Figure 8 shows, for example, that the evaluation result of the urban function "Super" is represented as a route length of "89.87". Furthermore, it shows that the parameters d(p,f) and d(f,v), used in the calculation formula for a predetermined route length described later, are represented as "16.35" and "73.52", respectively.

Figure 9 shows the initial setup screen for configuring the graph construction unit A1122. The settings screen W10 includes a screen W11 for selecting the target prefecture, a screen W12 for selecting the value information to be introduced, and a screen W13 for previewing the graph according to the selections made on screens W11 and W12. System users can use this screen to select one target prefecture and one or more value information items, and can also check a sample graph generated according to the settings on screens W11 and W12 on screen W13. The information displayed on screen W13 can be displayed by performing the processing S130 to S150 shown in Figure 12, which will be described later. In this example, when value information such as "ease of raising children," "ease of walking around," and "liveliness" is selected for an area belonging to the prefecture "Kanagawa Prefecture," the value information DBA1111 shown in Figure 3, the implemented urban function DBA1112 shown in Figure 4, and the spatial structure DBA1113 shown in Figure 5 are read, and a graph like the one described later is constructed, thereby creating a single graph that expresses the connection relationships of all "locations," "urban functions," and "value indicators." Furthermore, Figure 9 shows that the value information "Ease of raising children" and the area name "XX Shopping Street" have a high relationship between the nodes, resulting in a higher weight value of "30" being assigned to them compared to other nodes.

Traditionally, while it was possible to construct graphs representing the relationship between "urban functions" and "value indicators," location-specific information was missing, resulting in the same urban functions being recommended for many locations. However, in this embodiment, the map graph is interpreted as a graph where "space" and "urban functions" are fixed and value information is not included. The graph then focuses on the dependency characteristics of value indicators and urban functions mediated by urban space. In other words, by considering that "urban functions" and "value information" mutually influence each other through space, it becomes possible to separate "space" and "urban functions" and further realize a graph representation that links "value information" to "space."

Figure 10 shows the recommendation location setting screen for configuring the recommendation result generation unit A1124. The setting screen W20 includes a screen W21 for selecting the target location for recommendation, a screen W22 for setting the value indicators to be evaluated during recommendation and the importance of each value indicator, and a screen W23 for visualizing the relationships between each information item based on the settings in screens W21 and W22. System users can use this screen to select one target location for recommendation and one or more value information items, and can also set the importance of the selected value indicators numerically. Furthermore, the interrelationships between locations, urban functions, and value indicators, visualized based on the settings in screens W11 and W12, can be confirmed on screen W23. The information displayed on screen W23 can be displayed by performing the processing shown in S170 in Figure 12, which will be described later.

Figure 11 shows the results display screen for visualizing the output of the recommendation result generation unit A1124. The settings screen W30 contains screen W31 for displaying urban function recommendation results and screen W32 for visualizing the relationship with surrounding urban functions. System users can use this screen to check urban function recommendation results for the location set in the settings screen W20, and also to check the visualization of the relationship with surrounding urban functions when urban functions are introduced to the target location. The information displayed on screen W32 can be displayed by performing the processing S180 shown in Figure 12, which will be described later.

Referring to Figure 12, the operation for recommending urban functions in Example 1 will be described.

Step S110 is an operation in which the system accepts input of target prefectures and introduction value information from the system user in order to determine the database read range. The graph construction unit A1122 reads the target prefecture information and target value information from screens W11 and W12 transmitted from the user terminal A1200 via the user I/FA 1125.

Step S120 is the operation of searching for and reading records of the target prefecture from multiple databases in order to obtain the data necessary for recommendation generation from the database. The graph construction unit A1122 searches for and reads records in the value information DBA1111 where the value information name T10 matches the target value information and the prefecture T14 matches the target prefecture information. Next, it searches for records in the implementing city function DBA1112 where the prefecture T23 matches the target prefecture information, and records in the spatial structure DBA1113 where the prefecture T34 matches the target prefecture information, and reads the corresponding records.

Step S130 is the operation in which the graph construction unit A1122 determines the nodes and edges that make up the graph in order to construct a graph that represents the interrelationships between locations, urban functions, and value information. This step consists of multiple operations starting from step S131.

Step S131 is the operation for creating the nodes that make up the graph. Using the records read in Step S120 from the Value Information DBA1111, the Implemented Urban Function DBA1112, and the Spatial Structure DBA1113, the information items, node name, and node attributes that constitute the nodes are determined and registered in the node table of the Graph Structure DBA1114.

Specifically, records read from Value Function DBA1111 are converted into nodes with Value Information Name T10 as the node name, and the value of "Value Information" is set as the node attribute. Records read from Value Function DBA1111 are converted into nodes with Area IDT13 as the node name, and "Area" is set as the node attribute. Similarly, records read from Implementation City Function DBA1112 are converted into nodes with Location IDT20 as the node name, and "Location" is set as the node attribute. Next, a list of values set in Implementation City Function T22 is created from the records read from Implementation City Function DBA1112, and duplicate elements are removed from the list. Then, the elements in the list are converted into nodes, and "City Function" is set as the node attribute. The node name and node attribute information created through these operations is registered in Graph Structure DBA1114, a unique node IDT41 is assigned to all records in Graph Structure DBA1114, and step S131 is terminated.

Step S132 is the operation of setting an edge between nodes whose node attributes are "Value Information" and "Area". Referring to the value information DBA1111 read in step S120, the node pointed to by the value information name T10 and area IDT13 of each record is identified from the node table of the graph structure DBA1114, and the operation of setting an edge between the two identified nodes is repeated. Next, the edge configuration nodes identified by the above operation are registered in the edge table of the graph structure DBA1114, and step S132 ends.

Step S133 is the operation of setting weight information for the edges set in step S132. If information regarding the strength of the connection between the edges set in step S132 is available, this information is set as the edge weight. For example, if the value of the score T12 mentioned above is above a predetermined threshold, it can be determined that the relationship between the nodes is strong, and weight information can be set. Edges for which information regarding the strength of the connection is unavailable may not have an edge weight set.

Step S134 is the operation of setting an edge between nodes with the node attribute "Area" and "Location". Referring to the spatial structure DBA1113 read in step S120, for all records, the node pointed to by the location IDT31 and area IDT13 of the record is identified from the node table of the graph structure DBA1114, and the operation of setting an edge between the two identified nodes is repeated. Next, the edge configuration nodes identified by the above operation are registered in the edge table of the graph structure DBA1114, and step S134 ends.

Step S135 is the operation of setting edges between nodes whose node attribute is "Location". The record of the implemented city function DBA1112 read in step S120 is used as input. The adjacentness between nodes with the "Location" attribute is determined from the coordinate T21 contained in the record, and edges are set between nodes with adjacent relationships. This process of setting edges from coordinates can be performed using general methods such as a Nearest Neighbor Graph.

Step S136 is the operation of setting an edge between nodes with the node attribute "Location" and "City Function". Referring to the implemented city function DBA1112 read in step S120, for all records, the node pointed to by the location IDT20 and implemented city function T22 of the record is identified from the node table of the graph structure DBA1114, and the operation of setting an edge between the two identified nodes is repeated. Next, the edge configuration nodes identified by the above operation are registered in the edge table of the graph structure DBA1114, and step S136 ends.

Steps S140 and S150 are operations performed by the graph learning unit A1123 to obtain distributed representations of nodes from the graph by applying a graph embedding technique, which is one of the machine learning methods. Methods for obtaining distributed representations from a graph structure can include, for example, the method disclosed in Grover and Leskovec, “node2vec: Scalable Feature Learning for Networks”, Proceedings of the 22nd ACM SIGKDD international conference on Knowledge discovery and data mining, 2016.

A graph embedding method generally enables the extraction of distributed representations from a graph by performing two processes in this order: sampling graph nodes according to a certain rule, and calculating node distributed representations from the sampling results. In the embodiment of this disclosure, graph information is read from the graph structure DBA1114, graph nodes are sampled in step S140, and node distributed representations are calculated from the sampling results in step S141. Subsequently, the calculated node distributed representations are recorded in distributed representation A1115.

Step S160 is an operation to receive input for the recommended locations and the value information to be introduced. The recommendation result generation unit A1124 reads the setting information of screens W21 and W22 transmitted from the user terminal A1200 via the user I/FA 1125.

Step S170 is the operation in which the recommendation result generation unit A1124 calculates evaluation values at recommendation locations for urban functions that are not registered in the recommendation result DBA1116 among the candidate urban functions for recommendation. In this disclosure, the distributed representation vector of nodes recorded in the distributed representation DBA1115 is taken as input, and the path length corresponding to the distance traveled when traversing multiple nodes in the vector space is calculated and used as the evaluation value of the urban function.

The formula for calculating the path length in this disclosure is, for example, the following formula (1).

However, d(x,y) is a function representing the length of the line segment connecting the distributed representation vector of node x and the distributed representation vector of node y in the vector space, where p is any node whose node attribute is "location", f is any node whose node attribute is "urban function", and v is any node whose node attribute is "value information".

Equation (1) takes arbitrary nodes representing locations, urban functions, and value information as input and calculates the path length L when moving from a node with the "location" attribute, through a node with the "urban function" attribute, to a node representing "value information".

The first term of equation (1) represents the degree of affinity between the target location and urban functions; a smaller value indicates a higher affinity.

The second term of equation (1) represents the strength of the relationship between urban functions and value information; a smaller value indicates a stronger relationship.

The λ included in the second term of equation (1) is an adjustment parameter, and its value can be determined by the indicator importance on screen W22. This adjustment parameter allows for adjustment of the weighting of the first term, which indicates the degree of affinity between the target location and urban functions, and the second term, which indicates the strength of the relationship between urban functions and value information.

By using equation (1), it is possible to calculate an evaluation value for any urban function among the recommendation candidates, simultaneously considering its affinity with the function of the recommended location and its relationship with the value index.

Equation (1) considers only one evaluation indicator, but it is possible to consider multiple value indicators. In that case, equation (1) can be replaced with equation (2).

However, n is the number of value indicators to consider, and _vi is the distributed representation vector of the i-th value information node (where i is a natural number less than or equal to n).

The second term of equation (2) represents the sum of the strengths of the relationships between urban functions and all value information considered. _λi is an adjustment parameter for the i-th value information, and the weighting of importance between each piece of value information can be adjusted by changing the value of this parameter.

The urban functions, the calculated route length, and the values of d(p,f) and d(f,v) are recorded in the recommendation result DBA1116, and step S170 is terminated.

Step S171 is the operation in which the recommendation result generation unit A1124 determines whether the route length calculation shown in step S170 has been completed for all city functions that are candidates for recommendation. If the route length for all city functions is recorded in the recommendation result DBA1116 (step S171; Yes), the process proceeds to step S180; otherwise (step S171; No), it returns to step S170.

Step S180 is the operation in which the recommendation result generation unit A1124 takes the recommendation result DBA1116 as input, sorts urban functions in order of shortest route length, and transmits the results to the user terminal A1200 via the user I/FA1125. The user terminal A1200 presents the recommendation results to the system user using screen W31 within the result display screen W30. For example, for point 1 selected on screen W21 in Figure 10, the route length L is presented as the overall match degree, d(p,f) as the current match, and d(f,v) as the value match.

At this time, the recommendation result generation unit A1124 can also calculate the relationships between nodes of the "urban function" attribute recorded in the distributed representation DBA1115 using the distance between nodes, and visualize this on a map in screen W32 within the result display screen W30.

To facilitate user convenience, in step S110, a preview of the graph, which is immediately generated based on the selections made on screens W11 and W12, can be displayed on screen W13.

Furthermore, in step S160, depending on the selections made on screen W12, steps S170 and S171 can be executed immediately to display a visualization of the distributed representation vectors and path lengths between locations, urban functions, and value information on screen W23. In this case, to visualize higher-order distributed representation vectors on the screen, a dimensionality reduction algorithm such as t-SNE (t-Distributed Stochastic Neighbor Embedding), disclosed in Van der Maaten and Hinton, “Visualizing data using t-SNE”, Journal of machine learning research, 2008, can be used.

In Example 1, graph construction was performed based on database information. However, it is also possible that system users possess knowledge that is not included in the database but can be used for graph construction.

This embodiment describes an example of a recommendation generation system that can reflect the knowledge held by the system user as a foresight knowledge edge and can automatically determine the weight information of the foresight knowledge edge.

Regarding the recommendation generation system A1100 shown in Figure 1, the parts that have the same function as those already described in Figure 1 and are denoted by the same reference numerals will not be explained further.

Figure 13 shows the initial setup screen for configuring the graph construction unit A1122, which was done in Figure 9 in Example 1, in Example 2. The configuration screen W40 includes, in addition to the screens in the configuration screen W10 described in Figure 9, a screen W43 for accepting the addition of the system user's foresight knowledge edge.

Screen W43 includes a settings area W431, a checkbox W432, an input area W433, and a button W434.

The configuration area W431 accepts the selection of a combination of nodes to configure the foresight edge, enabling the addition of foresight as an edge to the graph.

Checkbox W432 is used to determine whether the recommendation generation system A1100 automatically sets the weight information for the edges configured in setting area W431. If the system user wishes to manually set the edge weight information using checkbox W432, they can input the edge weight information using input area W433.

After setting various information on screen W43, pressing button W434 sends the setting information to graph construction unit A1122. Graph construction unit A1122 registers the various information set on screen W43 in the edge table of graph structure DBA1114.
System users can view a sample graph, including the edge 1301 added according to the settings information on screen W43, on screen W44.

Referring to Figure 14, the operation for recommending urban functions in Example 2 will be described.

In Example 2, the same operations as in Example 1 are performed, except for steps S237, S2371, S2372, and S2373 described below.

Step S237 is an operation in which the graph construction unit A1122 determines whether the system user's foresight has been added. If the system has received input of foresight from the user on screen W43 (Step S237; Yes), it proceeds to step S2371. Otherwise (Step S237; No), it terminates the series of processes included in step S230.

Step S2371 is an operation in which the graph construction unit A1122 reflects the system user's foresight, which is not included in the database, into the graph. Based on the foresight information regarding the relationships between specific nodes received from the system user via screen W43, the unit identifies the relevant nodes from the graph structure DBA1114 and sets an edge between them. Next, the edge and its constituent nodes are registered in the edge table of the graph structure DBA1114, and step S238 is completed.

Step S2372 is an operation in which the graph construction unit A1122 determines whether the system user wishes to automatically set the weight information of the edges added in step S2371. If the user wishes to automatically set the weight information using the checkbox W432 on screen W43 (step S2372; Yes), the process proceeds to step S2373. Otherwise (step S2372; No), the series of processes included in step S230 is terminated.

Step S2373 is the operation by the graph construction unit A1122 to automatically set the weight information for the edges added in step S2371. In the automatic setting of edge weight information, first, the mutual reachability probability between all nodes is calculated from the graph constructed up to step S236. The mutual reachability probability between nodes can be solved by considering a Markov process where the nodes of the graph are states and the edge weights are transition probabilities, and applying a general method for deriving the stationary state of a Markov process.

Next, the attributes of the two nodes constituting the edge added in step S2371 are obtained (for example, "urban function" and "value information"), and all combinations of nodes with those attributes are enumerated (for example, a combination of a node with the "urban function" attribute and a node with the "value information" attribute). Furthermore, from the node combinations enumerated by the above operation, the node combination with the highest probability of reaching each other is identified by referencing the mutual reachability between the nodes that was solved. The weight information of the foresight knowledge edge is set so that the mutual reachability between the nodes in the identified node combination is equal to the mutual reachability between the nodes constituting the foresight knowledge edge.

The operation in step S2373 ensures, through mutual reachability probabilities, that the inter-node relationship in which foresight exists is the most important compared to other comparable inter-node relationships.

In Example 1, graph sampling was performed on the entire graph. However, since the graph constructed in Example 1 is expected to have nodes with the "Area" and "Urban Function" attributes connected to a large number of nodes with the "Location" attribute, it is conceivable that extracting relationships between nodes with the "Location" attribute will be difficult during sampling.

This embodiment describes an example of a recommendation generation system that can perform priority sampling of nodes with a specific attribute (e.g., the "location" attribute) when there is concern that the relationships between nodes of the same attribute may not be sufficiently estimated by machine learning methods due to tight coupling with numerous nodes of other attributes.

Regarding the recommendation generation system A1100 shown in Figure 1, the parts that have the same function as those already described in Figure 1 and are denoted by the same reference numerals will not be explained further.

Figure 15 shows the initial setup screen for configuring the graph construction unit A1122, which was done in Figure 9 in Example 1, in Example 3. The configuration screen W50 includes, in addition to the screens described in Figure 9, a screen W53 for receiving requests for priority sampling of nodes with specific attributes of the system user. Screen W53 includes a checkbox W531, a selection area W532, and an input area W533.

Checkbox W531 is a checkbox for system users to set whether or not to perform priority sampling. If a system user wishes to perform priority sampling using checkbox W531, they can input priority sampling settings using selection area W532 and input area W533.

System users can view a sample graph on screen W54, where the areas where priority sampling is performed are highlighted 1501 according to the settings information on screen W53.

Referring to Figure 16, the operation for recommending urban functions in Example 3 will be described.

In Example 3, the same operations as in Example 1 are performed, except for steps S350, S351, and S360, which are described below.

Step S350 is the operation in which the graph learning unit A1123 determines whether the user has requested that sampling be focused on nodes with specific attributes. If the user has made a setting using the checkbox W531 on screen W53 (Step S350; Yes), the process proceeds to step S351; otherwise (Step S350; No), the process proceeds to step S360.

Step S351 is the operation in which the graph learning unit A1123 constructs a subgraph consisting only of nodes with specific attributes and performs sampling from the subgraph. At this time, the number of samples to be taken from the subgraph can be determined from a ratio specified by the system user in the input area W533.

Step S360 is the operation in which the graph learning unit A1123 calculates the node distributed representation from the sampling results. In Example 1, the node distributed representation was calculated using only the sampling results for the entire graph. However, in this example, both the sampling results for the entire graph (step S340) and the sampling results for a subgraph consisting only of nodes with specific attributes (step S351) are used as training data to calculate the node distributed representation.

The embodiments have been described above, but in Embodiment 1, as explained with reference to Figure 12, etc., in a recommendation generation system A1100 that presents urban functions for a location using a computer having a processor and memory, the processor uses a first table (value information DBA1111) that associates first location information (points) with value information, and a second table (implemented urban functions DBA1112) that associates second location information (areas) containing at least one of a plurality of first location information with urban functions to set up graph nodes that point to the first location information and the second location information, as shown in Figures 6A and 6B. As shown, the process involves: a graph construction process that constructs a single graph representing the above urban functions as nodes connected to the first location information and the above value information as nodes connected to the second location information; a graph learning process that represents the interrelationships between nodes in a vector space using machine learning techniques based on the above graph; and a recommendation result generation process that evaluates the path length from the node-distributed representation vectors of the first location information, the above urban functions, and the above value information based on the specification of the first location information (location 1), evaluates the path length between nodes from the vector space, and generates recommendations for the above urban functions for the first location information. This allows for the automatic estimation of the range and extent of surrounding urban functions that may have affinity with the recommended location, and the explicit adjustment of the weighting of both affinity with surrounding urban functions and the relationship with value information when generating recommendations for urban functions for a given location.

Furthermore, in Example 2, as explained using Figure 14, the processor, in the graph construction process, adds foresight knowledge edges (edges 1301) between the nodes of the graph based on the user's specified foresight knowledge. The weight information of these added foresight knowledge edges is then set by comparing the reach probabilities between the graph nodes. This allows expert insights, which cannot be obtained from data alone, to be reflected in the graph.

Furthermore, in Example 3, as explained using Figure 16, the processor, in the graph learning process, extracts nodes with specific attributes (locations) from the graph as subgraphs, and generates learning data from the extracted subgraphs. This process generates learning data from both the graph and the subgraphs. As a result, even when area nodes and function nodes are connected to a large number of location nodes, the relationships between location nodes can be easily extracted.

The present invention is not limited to the embodiments described above. During implementation, the components can be modified and implemented without departing from the spirit of the invention, or multiple components disclosed in the embodiments can be appropriately combined.

A1100…Recommendation generation system, A1200…User terminal, A1300…Network, A1111…Value information DB, A1112…Implemented city function DB, A1113…Spatial structure DB, A1114…Graph structure DB, A1115…Distributed representation DB, A1116…Recommendation result DB, A1121…Automatic data update unit, A1122…Graph construction unit, A1123…Graph learning unit, A1124…Recommendation result generation unit, A1125…User interface

Claims (7)

A recommendation generation system that uses a computer having a processor and memory to present urban functions for a given location,
The aforementioned processor executes the program,
A graph construction process that uses a first table that associates areas , value information, and the prefecture to which the area belongs , and a second table that associates points included in at least one of multiple areas , urban functions, and the prefecture , to set up graph nodes that point to the areas and points, and constructs a single graph in which the urban functions are represented as nodes that connect to the points and the value information as nodes that connect to the areas .
Based on the aforementioned graph, a graph learning process is performed to represent the interrelationships between nodes in a vector space using a machine learning method,
Based on the designation of the aforementioned locations , the path length is evaluated from the node-distributed representation vector of the aforementioned locations , urban functions, and value information, thereby evaluating the path length between nodes from the vector space, and a recommendation result generation process is performed to generate recommendations for urban functions for the aforementioned locations .
A recommendation generation system characterized by performing the following. In the graph construction process, the processor adds foresight knowledge edges between the nodes of the graph based on the user's specified addition of foresight knowledge, and sets the weight information of the added foresight knowledge edges by comparing the reach probabilities between the graph nodes.
The recommendation generation system according to claim 1, characterized in that The processor generates training data from both the graph and the subgraphs by extracting nodes with specific attributes from the graph as subgraphs in the graph learning process, and by executing a process to generate training data from the extracted subgraphs.
The recommendation generation system according to claim 1, characterized in that A recommendation generation method that uses a computer having a processor and memory to present urban functions for a given location,
Using a first table that associates areas , value information, and the prefectures to which the areas belong , and a second table that associates locations included in at least one of multiple areas , urban functions, and the prefectures , a single graph is constructed in which nodes representing the areas and locations are set up, and the urban functions are represented as nodes connecting to the locations , and the value information as nodes connecting to the areas .
Based on the aforementioned graph, the interrelationships between nodes are represented in a vector space using a machine learning method.
Based on the designation of the aforementioned locations , the path length is evaluated from the node-distributed representation vector of the aforementioned locations , the aforementioned urban functions, and the aforementioned value information, thereby evaluating the path length between nodes from the vector space, and a recommendation for the aforementioned urban functions is generated for the aforementioned locations .
A recommendation generation method characterized by the following. In constructing the aforementioned graph, based on the user's specified addition of foresightful knowledge, foresightful knowledge edges are added between the nodes of the graph, and the weight information of these added foresightful knowledge edges is set by comparing the reach probabilities between the graph nodes.
The recommendation generation method according to claim 4, characterized in that In the learning process of the graph, nodes with specific attributes are extracted from the graph as subgraphs, and training data is generated from the extracted subgraphs, thereby generating training data from both the graph and the subgraphs.
The recommendation generation method according to claim 4, characterized in that A computer having a processor and memory,
A graph construction process that uses a first table that associates areas , value information, and the prefecture to which the area belongs , and a second table that associates points included in at least one of multiple areas , urban functions, and the prefecture , to set up graph nodes that point to the areas and points, and constructs a single graph in which the urban functions are represented as nodes that connect to the points and the value information as nodes that connect to the areas .
Based on the aforementioned graph, a graph learning process is performed to represent the interrelationships between nodes in a vector space using a machine learning method,
Based on the designation of the aforementioned locations , the path length is evaluated from the node-distributed representation vector of the aforementioned locations , urban functions, and value information, thereby evaluating the path length between nodes from the vector space, and a recommendation result generation process is performed to generate recommendations for urban functions for the aforementioned locations .
A recommendation generation program characterized by executing [the specified action].