Method and device for acquiring representation information of point of interest (POI) and electronic equipment

Abstract

The present disclosure provides a point of interest (POI) representation information acquisition method and device, and electronic equipment, relates to the technical field of data processing, and particularly relates to the fields of artificial intelligence, deep learning, information mining, and autonomous driving. The specific implementation scheme is as follows: a POI relationship graph is acquired, and based on the POI relationship graph, a node relationship set of each node on the POI relationship graph is acquired, wherein the POI relationship graph includes nodes representing POIs and edges representing relationships between nodes, and the node relationship set includes a relationship path of the node reaching all second-order neighbor nodes via any first-order neighbor node; a spatial distance matrix between the node and the corresponding second-order neighbor nodes is acquired; and based on the node relationship set and the spatial distance matrix of the node, first POI representation information of the node is determined. The present disclosure fully mines relationship information and spatial information in the second-order neighborhood of the node, and the acquired first POI representation information is more accurate and representative.

Description

Technical Field

[0001] This disclosure relates to the field of data processing technology, specifically to the fields of artificial intelligence, deep learning, information mining, and autonomous driving, and particularly to a method, apparatus, and electronic device for acquiring representational information of points of interest (POIs). Background Technology

[0002] Points of interest (POIs) on a map are a basic necessity for residents' daily travel and have significant commercial value. They can provide cities and users with comprehensive location information, facilitating better management and planning of travel needs.

[0003] Currently, when determining the learning and acquisition of POI representation information, the relative distances of neighboring nodes are directly aggregated, which easily leads to information loss and ignores the complex relationships between POIs, resulting in insufficient mining of POI representation information. Summary of the Invention

[0004] This disclosure provides a method, apparatus, and electronic device for obtaining representational information of points of interest (POIs).

[0005] According to one aspect of this disclosure, a method for obtaining representation information of Points of Interest (POIs) is provided. The method includes: obtaining a POI relationship graph, and based on the POI relationship graph, obtaining a set of node relationships for each node in the POI relationship graph, wherein the POI relationship graph includes nodes representing POIs and edges representing relationships between nodes, and the set of node relationships includes relational paths from a node to all second-order neighbor nodes via any first-order neighbor node; obtaining a spatial distance matrix between the node and its corresponding second-order neighbor nodes; and determining first POI representation information of the node based on the set of node relationships and the spatial distance matrix.

[0006] According to another aspect of this disclosure, a training method for a spatial relationship-aware model of Points of Interest (POIs) is provided. The method includes: acquiring a sample POI relationship graph, and based on the sample POI relationship graph, acquiring a set of sample node relationships for each sample node in the sample POI relationship graph, wherein the sample POI relationship graph includes sample nodes representing POIs and edges representing relationships between sample nodes, and the set of sample node relationships includes relationship paths from a sample node to all second-order neighbor nodes via any first-order neighbor node; acquiring a sample spatial distance matrix between the sample node and its corresponding second-order neighbor node; training an initial spatial relationship-aware model based on the set of sample node relationships and the sample spatial distance matrix to determine first sample POI representation information for the sample node; obtaining downstream task prediction data for the sample node based on the first sample POI representation information, and correcting the spatial relationship-aware model based on the downstream task prediction data and the downstream task annotation data for the sample node, until training is completed and a target spatial relationship-aware model is obtained.

[0007] According to a third aspect of this disclosure, an apparatus for acquiring representation information of Points of Interest (POIs) is provided, comprising: a first acquisition module, configured to acquire a POI relationship graph and, based on the POI relationship graph, acquire a set of node relationships for each node in the POI relationship graph, wherein the POI relationship graph includes nodes representing POIs and edges relating to the relationships between nodes, and the set of node relationships includes relational paths from a node to all second-order neighbor nodes via any first-order neighbor node; a second acquisition module, configured to acquire a spatial distance matrix between the node and its corresponding second-order neighbor nodes; and an information acquisition module, configured to determine first POI representation information of the node based on the set of node relationships and the spatial distance matrix.

[0008] According to a fourth aspect of this disclosure, a training apparatus for a spatial relationship awareness model of Points of Interest (POIs) is provided, comprising: a first sample acquisition module, configured to acquire a sample POI relationship graph, and based on the sample POI relationship graph, acquire a set of sample node relationships for each sample node on the sample POI relationship graph, wherein the sample POI relationship graph includes sample nodes representing POIs and edges representing relationships between sample nodes, and the set of sample node relationships includes relationship paths from a sample node to all second-order neighbor nodes via any first-order neighbor node; a second sample acquisition module, configured to acquire a sample spatial distance matrix between the sample node and its corresponding second-order neighbor node; a sample information acquisition module, configured to train an initial spatial relationship awareness model based on the set of sample node relationships and the sample spatial distance matrix, and determine first sample POI representation information of the sample node; and a correction module, configured to obtain downstream task prediction data of the sample node based on the first sample POI representation information, and correct the spatial relationship awareness model based on the downstream task prediction data and the downstream task annotation data of the sample node, until training is completed and a target spatial relationship awareness model is obtained.

[0009] According to a fifth aspect of this disclosure, an electronic device is provided, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the method for obtaining representational information of points of interest (POIs) according to a first aspect of this disclosure or the method for training a spatial relationship-aware model of POIs according to a second aspect of this disclosure.

[0010] According to a sixth aspect of this disclosure, a non-transitory computer-readable storage medium storing computer instructions is provided, wherein the computer instructions are configured to cause the computer to execute a method for obtaining representational information of points of interest (POIs) according to a first aspect of this disclosure or a method for training a spatial relationship-aware model of POIs according to a second aspect of this disclosure.

[0011] According to a seventh aspect of this disclosure, a computer program product is provided, comprising a computer program that, when executed by a processor, implements a method for obtaining representational information of points of interest (POIs) according to a first aspect of this disclosure or a method for training a spatial relationship-aware model of POIs according to a second aspect of this disclosure.

[0012] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of this disclosure, nor is it intended to limit the scope of this disclosure. Other features of this disclosure will become readily apparent from the following description. Attached Figure Description

[0013] The accompanying drawings are provided to better understand this solution and do not constitute a limitation of this disclosure. Wherein:

[0014] Figure 1 This is a schematic diagram of a method for obtaining representation information of a point of interest (POI) provided in an embodiment of this disclosure;

[0015] Figure 1a This is a schematic diagram of node relationships provided in an embodiment of this disclosure;

[0016] Figure 2 This is a schematic diagram of another method for obtaining representation information of a point of interest (POI) provided in an embodiment of this disclosure;

[0017] Figure 2a This is a schematic diagram of the spatial distance between nodes provided in an embodiment of this disclosure;

[0018] Figure 3 This is a schematic diagram of another method for obtaining representation information of a point of interest (POI) provided in an embodiment of this disclosure;

[0019] Figure 4 This is a schematic diagram of another method for obtaining representation information of a point of interest (POI) provided in an embodiment of this disclosure;

[0020] Figure 4a This is a logical schematic diagram of another method for obtaining representation information of Point of Interest (POI) provided in this embodiment of the disclosure;

[0021] Figure 5 This is a schematic diagram of a training method for a spatial relationship-aware model of points of interest (POIs) provided in an embodiment of this disclosure;

[0022] Figure 6 This is a schematic diagram of another method for obtaining representation information of a point of interest (POI) provided in an embodiment of this disclosure;

[0023] Figure 7 This is a structural block diagram of a device for acquiring representational information of Point of Interest (POI) provided in an embodiment of this disclosure;

[0024] Figure 8 This is a structural block diagram of a training device for a spatial relationship perception model of points of interest (POIs) provided in an embodiment of this disclosure;

[0025] Figure 9 This is a block diagram of an electronic device used to implement a method for obtaining representational information of a point of interest (POI) or a method for training a spatial relationship perception model of a POI provided in the embodiments of this disclosure. Detailed Implementation

[0026] The exemplary embodiments of this disclosure are described below with reference to the accompanying drawings, including various details of the embodiments to aid understanding, and should be considered merely exemplary. Therefore, those skilled in the art will recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of this disclosure. Similarly, for clarity and brevity, descriptions of well-known functions and structures are omitted in the following description.

[0027] Data processing is the process of collecting, storing, retrieving, processing, transforming, and transmitting data. It involves extracting and deriving valuable and meaningful data from large amounts of potentially chaotic and incomprehensible data.

[0028] Artificial Intelligence (AI) is a new technical science that studies, develops, and applies theories, methods, technologies, and application systems to simulate, extend, and expand human intelligence. It attempts to understand the nature of intelligence and produce new intelligent machines that can react in a way similar to human intelligence. Research in this field includes robotics, speech recognition, image recognition, natural language processing, and expert systems.

[0029] Deep learning is a new research direction in the field of machine learning. Deep learning learns the inherent patterns and hierarchical representations of sample data. The information gained during this learning process greatly aids in the interpretation of data such as text, images, and sound. Its ultimate goal is to enable machines to possess analytical and learning capabilities like humans, and to recognize data such as text, images, and sound.

[0030] Information mining is the process of extracting intrinsic characteristics from a large number of samples to conduct targeted information extraction. It utilizes information technology to analyze information resources and extract previously unknown, complete information from various information sources for critical business decision-making. It includes two types: data mining and text mining.

[0031] Autonomous driving generally refers to automatic driving systems that employ advanced communication, computer, network, and control technologies to achieve real-time, continuous control of trains. Utilizing modern communication methods and directly interacting with the train, it enables two-way data communication between the train and the ground, resulting in high transmission rates and large information volumes. This allows subsequent train tracking and the control center to promptly obtain the precise location of the preceding train, making operation management more flexible, control more effective, and better suited to the needs of automatic train driving.

[0032] Figure 1 This is a schematic diagram of a method for obtaining representational information of Points of Interest (POIs) provided in an embodiment of this disclosure. Figure 1 As shown, the method includes, but is not limited to, the following steps:

[0033] S101, obtain the POI relationship graph, and based on the POI relationship graph, obtain the set of node relationships for each node in the POI relationship graph.

[0034] The POI relationship graph includes nodes representing POIs and edges representing relationships between nodes. The node relationship set includes the relationship paths from a node to all second-order neighbors via any first-order neighbor node.

[0035] First-order neighbors are nodes directly adjacent to a given node, while second-order neighbors are nodes connected via first-order neighbors. Second-order neighbors are adjacent to first-order neighbors. Figure 1a As shown, node V i The first-order neighbor node is node V j Node V i The second-order neighbor node is V k .

[0036] The purpose of this disclosure is to obtain accurate POI representation information for each POI node; more accurate POI representation information can be applied to more scenarios, such as recommending users' travel, capturing high-value customers for enterprises, and providing merchants with precise online promotion.

[0037] Optionally, a POI relationship graph can be determined based on an existing map. This POI relationship graph contains different nodes and the relationships between nodes. The nodes in the POI relationship graph represent different POIs. When there is a certain relationship between the nodes, the corresponding nodes will be connected. That is to say, the POI relationship graph includes nodes representing POIs and edges representing the relationships between nodes.

[0038] In some implementations, relationships are considered to exist between POI nodes when there are competition, complementarity, or dependence among them.

[0039] In other implementations, a relationship is considered to exist between POI nodes when they are frequently accessed by a common user. For example, if two POI nodes are accessed by a common user more than a certain number of times consecutively, it is assumed that these two POI nodes are frequently accessed by a common user, and thus a relationship is considered to exist between them.

[0040] Furthermore, the set of node relationships for each node in the POI relationship graph is determined; the set of node relationships includes all relationships between nodes, as well as all relationship paths. Optionally, a relationship path is the path from a node to all its second-order neighbor nodes via its first-order neighbor nodes.

[0041] For example, suppose the current node V i With node V j There is a competitive relationship between node V. j With node V k There is also competition between them, so we can infer that node V i With node V k There is also a high possibility that there is competition between them, then node V j As node V i First-order neighbor node, node V k As node V i Establish node V by identifying its second-order neighbor nodes. i To node V j To node V k The relationship path. The node relationship set includes the relationship paths corresponding to all nodes. The relationship set can be used to quickly determine the relationships between nodes, which is convenient for subsequent information mining based on the relationships between nodes.

[0042] S102, obtain the spatial distance matrix between the node and its corresponding second-order neighbor nodes.

[0043] Under normal circumstances, the distance between POI nodes can reflect the correlation between them; the closer the distance, the stronger the correlation. Therefore, in addition to considering the relationship between POI nodes, we also need to analyze the spatial distance between POI nodes.

[0044] Optionally, the spatial distance between a POI node and its corresponding second-order neighbor node can be obtained based on the coordinates of each node. The spatial distance can be Euclidean distance, Manhattan distance, or Chebyshev distance, etc.

[0045] Alternatively, the coordinates of each node can be obtained from the latitude and longitude on the map, or by reconstructing the coordinate system.

[0046] After obtaining the spatial distances between each node and all its corresponding second-order neighbor nodes, the spatial distances of all nodes are summarized to obtain a spatial distance matrix.

[0047] S103, based on the node relationship set and spatial distance matrix of the node, determine the first POI representation information of the node.

[0048] It is understandable that the node relationship set contains complex relationships between nodes, and the spatial distance matrix represents the spatial characteristic information between nodes. Therefore, the node relationship set and the spatial distance matrix can reflect the characteristic information corresponding to each node in a good way.

[0049] Optionally, features can be extracted from the node relationship set and the spatial distance matrix separately, and the extracted features can be fused to obtain the first POI representation information of the node. In this embodiment of the disclosure, a POI relationship graph and a node relationship set in the POI relationship graph are obtained. The node relationship set includes all complex relationships between nodes. The spatial distance matrix is ​​determined based on the spatial location of the nodes. The spatial distance matrix reflects the spatial characteristic information between nodes. The first POI representation information is determined based on the node relationship set and the spatial distance matrix. This overcomes the problem of missing distance information in the prior art, can better preserve the spatial characteristics of long distances, and fully explores the relationship information and spatial information within the second-order neighborhood of the node. The obtained first POI representation information is more accurate and representative.

[0050] Figure 2 This is a schematic diagram of a method for obtaining representational information of Points of Interest (POIs) provided in an embodiment of this disclosure. Figure 2 As shown, the method includes, but is not limited to, the following steps:

[0051] S201, obtain the POI relationship graph, and based on the POI relationship graph, obtain the set of node relationships for each node in the POI relationship graph.

[0052] The POI relationship graph includes nodes representing POIs and edges representing relationships between nodes. The node relationship set includes the relationship paths from a node to all second-order neighbors via any first-order neighbor node.

[0053] In some implementations, the set of node relationships can be obtained based on the relationship between a node and all its corresponding second-order neighbor nodes.

[0054] In other implementations, since the number of second-order neighbor nodes is quite large, but not all second-order neighbor nodes contain rich second-order neighborhood information, the second-order neighbor nodes of each node can be filtered to improve the accuracy of second-order neighborhood information mining.

[0055] Optionally, for node V i Any candidate second-order neighbor node V k Based on the POI relationship graph, obtain the slave node V. i After passing through any first-order neighbor node V j Arrival at node V k The number of paths in the relational path; from node V i Among all candidate second-order neighbor nodes, one or more candidate second-order neighbor nodes with a number of paths greater than a set number are identified and retained as the target second-order neighbor nodes of the node; the relationship paths between the node and the target second-order neighbor nodes are mined to obtain the node relationship set of the node.

[0056] For example, for node V i and its corresponding candidate second-order neighbor node V k Obtain node V in the POI relationship graph i After passing through any first-order neighbor node V j Arrived candidate second-order neighbor node V k The number of all relational paths between them; where node V i After passing through any first-order neighbor node V j Arrived candidate second-order neighbor node V k All relationship paths between them can be represented as: R i，j It is node V i Relationship with first-order neighbor nodes, R j，k It is a first-order neighbor node V j and second-order neighbor node V k The relationship; if the number of paths is greater than the set number, then the candidate second-order neighbor node V k For node V i The target second-order neighbor node; and so on, to determine node V. i The number of relationship paths to each candidate second-order neighbor node is used to determine node V based on whether the number of paths exceeds a set limit. i All corresponding target second-order neighbor nodes; for node V i The relationship paths between the target and all its corresponding second-order neighbor nodes are mined to obtain a set of node relationships.

[0057] Alternatively, the second-order neighborhood relation can also be represented as follows:

[0058]

[0059] in, Represents node V i The second-order neighborhood relation; V j Represents node V i First-order neighbor node; V k Represents node V i The second-order neighbor nodes; r1→r2 represents a relation schema, where r1 is a node V. i With first-order neighbor node V j The relationship is that r2 is a first-order neighbor node V. j and second-order neighbor node V k Relationship; R i，j It is node V i The relationship with first-order neighbor nodes is the same as that of r1; R j，k It is a first-order neighbor node V j and second-order neighbor node Vk The relationship is the same as that of r2.

[0060] In this embodiment of the disclosure, step S201 can be implemented in any of the various embodiments of the disclosure, and no limitation is made here, nor will it be described in detail.

[0061] S202, obtain the spatial distance matrix between the node and its corresponding second-order neighbor nodes.

[0062] Optionally, for any second-order neighbor node of a node, obtain the first spatial coordinate information of the node and the second spatial coordinate information of the second-order neighbor node; perform Euclidean distance calculation based on the first spatial coordinate information and the second spatial coordinate information to obtain the spatial distance between the node and the second-order neighbor node; and obtain the spatial distance matrix based on the spatial distance between the node and each second-order neighbor node.

[0063] like Figure 2a As shown, this diagram illustrates the spatial distance between nodes; for example, the spatial distance between a node and its second-order neighbor can be:

[0064] d ik =|L i -L k |,k∈(r1→r2)

[0065] Where, d ik L represents the spatial distance between node i and its second-order neighbor node k; i L represents the first spatial coordinate information of node i; k This represents the second spatial coordinate information of the second-order neighbor node k; r1→r2 represents a relational pattern, where r1 is the relationship between the node and its first-order neighbor node, and r2 is the relationship between the first-order neighbor node and its second-order neighbor node.

[0066] In some implementations, the first spatial coordinates of a node and the second spatial coordinates of its second-order neighbor nodes can be determined by latitude and longitude on a map. That is, the first spatial coordinates are the latitude and longitude of the node, and the second spatial coordinates are the latitude and longitude of its second-order neighbor nodes, which better preserves the spatial characteristics of the node.

[0067] Optionally, the spatial distances from a node to each of its second-order neighbor nodes can be arranged sequentially to obtain a spatial distance matrix. For example, if there are N nodes, and each node corresponds to M second-order neighbor nodes, then the spatial distances between each node and all its second-order neighbor nodes can form a submatrix of one row and M columns. By concatenating N submatrices of one row and M columns, an N-row and M-column spatial distance matrix can be obtained.

[0068] In this embodiment of the disclosure, step S202 can be implemented in any of the various embodiments of the disclosure, and no limitation is made here, nor will it be described in detail.

[0069] S203, determine the relation subsets of nodes under different relation schemas from the node relation set, wherein the relation subsets include at least one relation path of nodes under the same relation schema.

[0070] Because the set of node relationships contains various relationship patterns, such as competition, complementarity, and dependency, a subset of relationships under different relationship patterns is determined from the set of node relationships. Each subset of relationships includes at least one relationship path between nodes under the same relationship pattern.

[0071] For example, for a given node, the relationships that can exist with other nodes include competitive, complementary, or dependent relationships. Correspondingly, the relationship paths from this node to its second-order neighbors can include various relationship paths such as: "competition + competition," "competition + complementarity," "competition + dependency," "complementarity + competition," "complementarity + complementarity," "complementarity + dependency," "dependency + competition," "dependency + complementarity," and "dependency + dependency." The relationship paths of nodes with the same relationship pattern are aggregated into corresponding relationship subsets. For example, all "competition + competition" relationship paths corresponding to a node are aggregated to obtain the corresponding relationship subset.

[0072] S204. For each relation schema, based on the relation paths and spatial distance matrix in the relation subset, determine the second POI representation information of the node in the relation schema.

[0073] Since the relational path of a relation subset includes the first-order and second-order neighbor nodes of a node, the information represented by the first-order and second-order neighbor nodes of the node needs to be considered when determining the second POI representation information of a node.

[0074] Optionally, for each relation path in the relation subset corresponding to the same relation schema, the spatial distance between the node in the relation path and its second-order neighbor node is determined based on the spatial distance matrix; the third POI representation information of the node under the relation path is determined based on the spatial distance and the first-order and second-order neighbor nodes of the node in the relation path; and the third POI representation information of the node under the same relation schema is fused to obtain the second POI representation information of the node under the same relation schema.

[0075] In other words, for any relation path in the relation subset, the spatial distance between a node and its second-order neighbors can be determined based on the spatial distance matrix between the node and all its second-order neighbors. The smaller the distance, the stronger the correlation between the node and its second-order neighbors in the relation path.

[0076] Optionally, the relationship between a node's first-order neighbor nodes and second-order neighbor nodes can be determined based on the relational path. Based on the spatial distance between a node and its second-order neighbor nodes, and the relationship between a node's first-order neighbor nodes and second-order neighbor nodes in the relational path, the third POI representation information of the node under the relational path can be determined.

[0077] Optionally, the spatial distance between a node and its second-order neighbor nodes, and the relationship between the first-order and second-order neighbor nodes of a node in the relational path, can be input into the pre-trained model to output the third POI representation information.

[0078] Furthermore, after determining the third POI representation information of a node under each relation path, the third POI representation information of the node under the same relation schema is fused to obtain the second POI representation information of the node under the same relation schema. This fusion can be achieved by summing all the third POI representation information or by calculating the average information of all the third POI representation information. The second POI representation information of the node under the same relation schema is determined through the fusion of all the third POI representation information, making the features contained in the second POI representation information more comprehensive and accurate.

[0079] S205, perform a fusion operation on the second POI representation information of the node in each relation schema to obtain the first POI representation information of the node in all relation schemas.

[0080] Optionally, after determining the second POI representation information of a node in each relation schema, a fusion operation can be performed on the second POI representation information of the node in each relation schema to obtain the first POI representation information of the node in all relation schemas.

[0081] Optionally, the fusion operation can be to accumulate the second POI representation information of the node in each relation mode, or to calculate the average information of the second POI representation information of the node in all relation modes, and finally confirm the first POI representation information of the node in all relation modes, so as to ensure that the first POI feature information is more accurate.

[0082] For example, the method for obtaining the feature information of the first POI can be:

[0083]

[0084] Among them, h i,intra Indicates the first POI feature information; The second POI represents the information; |*R| represents the number of all relation schemas; r1→r2 represents a relation schema, where r1 is the relationship between a node and its first-order neighbor nodes, and r2 is the relationship between a first-order neighbor node and its second-order neighbor nodes.

[0085] In other words, by averaging the second POI representation information of all relation patterns, the first POI feature information of the node under all relation patterns is obtained.

[0086] Optionally, after determining the first POI feature information of a node in all relational modes, downstream task processing can be performed on the node based on the first POI feature information to obtain the processing results of the downstream task.

[0087] Understandably, downstream tasks are the actual tasks that need to be solved, and the processing result of a node's downstream task can be considered as whether that node can solve the downstream task. For example, in this embodiment, it is necessary to recommend user travel; after determining the first POI feature information of each node, the node's recommendation value can be obtained based on the node's first POI feature information. Based on the node's recommendation value, it can be determined whether the node needs to be recommended, that is, whether the node can be used as the final result of the downstream task, thus ensuring the accuracy of the downstream task.

[0088] In this embodiment, a subset of relationships of nodes under different relationship patterns is determined from the set of node relationships. By analyzing the subset of relationships under the same relationship pattern, the accuracy and reliability of the analysis are ensured. For each relationship pattern, the spatial distance between a node and its second-order neighbor node in each relationship path of the subset of relationships is obtained. Based on the spatial distance and the relationship between first-order and second-order neighbor nodes in the relationship path, third POI representation information is determined. The third POI representation information of each relationship path is fused to obtain second POI representation information, ensuring that the second POI representation information contains more comprehensive information. Furthermore, all second POI representation information is fused to obtain the first POI representation information of the node, which better preserves complex relationships and spatial distances and improves the accuracy of the first POI representation information. Downstream task processing is performed based on more accurate first POI representation information, and the downstream task processing results are also more accurate.

[0089] Figure 3 This is a schematic diagram of a method for obtaining representational information of Points of Interest (POIs) provided in an embodiment of this disclosure. Figure 3 As shown, the method includes, but is not limited to, the following steps:

[0090] S301, obtain the POI relationship graph, and based on the POI relationship graph, obtain the set of node relationships for each node in the POI relationship graph.

[0091] The POI relationship graph includes nodes representing POIs and edges representing relationships between nodes. The node relationship set includes the relationship paths from a node to all second-order neighbors via any first-order neighbor node.

[0092] In this embodiment of the disclosure, the method for implementing step S301 can be implemented in any of the various embodiments of the disclosure, and no limitation is made here, nor will it be described in detail.

[0093] S302, obtain the spatial distance matrix between the node and its corresponding second-order neighbor nodes.

[0094] In this embodiment of the disclosure, the method for implementing step S302 can be implemented in any of the various embodiments of the disclosure, and no limitation is made here, nor will it be described in detail.

[0095] S303, determine the relation subsets of nodes under different relation schemas from the node relation set, wherein the relation subsets include at least one relation path of nodes under the same relation schema.

[0096] In this embodiment of the disclosure, the method for implementing step S303 can be implemented in any of the various embodiments of the disclosure, and no limitation is made here, nor will it be described in detail.

[0097] S304: For each relation path in the relation subset corresponding to the same relation schema, determine the spatial distance between the node in the relation path and its second-order neighbor node based on the spatial distance matrix.

[0098] In this embodiment of the disclosure, the method for implementing step S304 can be implemented in any of the various embodiments of the disclosure, and no limitation is made here, nor will it be described in detail.

[0099] S305 determines the relation weight vector of a node under the relation path based on first-order neighbor nodes and second-order neighbor nodes.

[0100] Optionally, determine the first feature vector of the first-order neighbor node and the second feature vector of the second-order neighbor node of the node in the relation path; perform a linear transformation on the first feature vector and the second feature vector to determine the relation weight vector corresponding to the node.

[0101] Optionally, the first feature vector of the first-order neighbor node and the second feature vector of the second-order neighbor node can be feature vectors extracted based on information such as the POI category of the neighbor node, used to characterize the features of the first-order neighbor node and the second-order neighbor node. Optionally, the first feature vector and the second feature vector can have the same dimension or different dimensions.

[0102] Optionally, the linear transformation can be splicing, projection, scaling, or rotation, etc.

[0103] S306 determines the spatial weight vector of a node under the relation path based on spatial distance and second-order neighbor nodes.

[0104] Optionally, the spatial distance representation vector and the second feature vector of the second-order neighbor node are determined; the spatial distance representation vector and the second feature vector are linearly transformed to determine the spatial weight vector corresponding to the node.

[0105] Optionally, the spatial distance representation vector is used to reflect the characteristics of the spatial distance. The spatial distance representation vector can have the same dimension as the first feature vector and the second feature vector.

[0106] Optionally, the linear transformation can be concatenation, projection, scaling, or rotation. For example, the spatial distance representation vector and the second feature vector of the second-order neighbor node can be concatenated, and the concatenated vector is determined as the spatial weight vector corresponding to the node.

[0107] S307, the spatial weight vector and relational weight vector are fused to obtain the weight coefficients of the second-order neighbor nodes.

[0108] Optionally, the spatial weight vector and relational weight vector can be fused using either a weighted average method or principal component analysis. The fused weight coefficients include both spatial and relational information between nodes. The fused weight coefficients are then used to obtain the third point of interest (POI) representation information of the nodes, which is more representative.

[0109] S308, based on first-order neighbor nodes, second-order neighbor nodes and weight coefficients, obtains the third POI representation information of the node under the relation path.

[0110] Optionally, the third POI representation information of a node under the relation path can be obtained by fusing the first feature vector of the first-order neighbor node, the second feature vector of the second-order neighbor node, and the weight coefficient.

[0111] Since the weight coefficient includes spatial distance information and relationship information between the node and its second-order neighbor nodes, it can be used as the weight of the second feature vector of the second-order neighbor node. The weight coefficient is used to perform a weighted operation on the second feature vector and then fused with the first feature vector to obtain the third POI representation information of the node under the relationship path.

[0112] Optionally, the third POI representation information can be obtained by adding the first feature vector to the vector obtained by weighting the second feature vector using weight coefficients.

[0113] S309, perform a fusion operation on all third POI representation information of the node under the same relation schema to obtain the second POI representation information of the node under the same relation schema.

[0114] In this embodiment of the disclosure, the method for implementing step S309 can be implemented in any of the various embodiments of the disclosure, and no limitation is made here, nor will it be described in detail.

[0115] S310, perform a fusion operation on the second POI representation information of the node in each relation schema to obtain the first POI representation information of the node in all relation schemas.

[0116] In this embodiment of the disclosure, the method for implementing step S310 can be implemented in any of the various embodiments of the disclosure, and no limitation is made here, nor will it be described in detail.

[0117] In this embodiment, a first feature vector of a first-order neighbor node and a second feature vector of a second-order neighbor node are determined to represent the features corresponding to the first-order and second-order neighbor nodes, respectively. A relation weight vector is obtained by performing linear transformations such as concatenation based on the first and second feature vectors. A representation vector representing the spatial distance feature is determined, and a spatial weight vector is determined using this representation vector and the second feature vector. The relation weight vector and the spatial weight vector are fused to obtain a weight coefficient, which includes the relational information between nodes and the spatial distance information. The second feature vector is weighted using the weight coefficient, and combined with the first feature vector to obtain the third POI representation information. By measuring the influence of the second-order neighbor node using both aspects of information, the third POI representation information is more accurate, and the first POI representation information obtained using the third POI representation information is more convincing and accurate.

[0118] Figure 4 This is a schematic diagram of a method for obtaining representational information of Points of Interest (POIs) provided in an embodiment of this disclosure. Figure 4 As shown, the method includes, but is not limited to, the following steps:

[0119] S401, obtain the POI relationship graph, and based on the POI relationship graph, obtain the set of node relationships for each node in the POI relationship graph.

[0120] The POI relationship graph includes nodes representing POIs and edges representing relationships between nodes. The node relationship set includes the relationship paths from a node to all second-order neighbors via any first-order neighbor node.

[0121] In this embodiment of the disclosure, the method for implementing step S401 can be implemented in any of the various embodiments of the disclosure, and no limitation is made here, nor will it be described in detail.

[0122] S402, obtain the spatial distance matrix between the node and its corresponding second-order neighbor nodes.

[0123] In this embodiment of the disclosure, the method for implementing step S402 can be implemented in any of the various embodiments of the disclosure, and no limitation is made here, nor will it be described in detail.

[0124] S403, Obtain a pre-trained target spatial relationship perception model.

[0125] Optionally, the representation information of the first POI can be acquired based on a target spatial relationship awareness model. The target spatial relationship awareness model can be a graph neural network to facilitate the extraction and discovery of features and patterns in graph structure data, thereby improving the efficiency and accuracy of acquiring the representation information of the first POI.

[0126] S404 uses the first target linear transformation matrix of the spatial relationship perception layer to perform linear transformation on the first feature vector of the first-order neighbor node and the second feature vector of the second-order neighbor node of the node, thereby determining the relationship weight vector of the node under the relationship path.

[0127] In some implementations, the spatial relationship perception layer is a processing layer of the target spatial perception model, and the spatial relationship perception layer corresponds to the first target linear transformation matrix, which can be obtained during the training process of the target spatial perception model.

[0128] Furthermore, based on the first objective linear transformation matrix, the first eigenvector of the first-order neighbor node and the second eigenvector of the second-order neighbor node of the node are linearly transformed to determine the relation weight vector of the node under the relation path; that is, the relation weight vector is obtained according to the first objective linear transformation matrix, the first eigenvector and the second eigenvector.

[0129] Alternatively, the relation weight vector can be obtained by:

[0130]

[0131] in, The vector represents the relation weights, where r1 and r2 represent the relations in the relation path, namely the relationship between a node and its first-order neighbor nodes, and the relationship between a first-order neighbor node and its second-order neighbor nodes, respectively. This represents the linear transformation matrix of the target relation, used for... The fusion and transformation can be performed again, which can be obtained during the training process of the target space perception model; h represents the first objective linear transformation matrix; j h represents the first eigenvector of a first-order neighbor node. k This represents the second feature vector of a second-order neighbor node; This represents the concatenation symbol, used to concatenate vectors.

[0132] In other words, the first target linear transformation matrix is ​​multiplied by the first eigenvector and the second eigenvector respectively, and the resulting vectors are concatenated to determine the relation weight vector. This relation weight vector integrates the self-features of the first-order neighbor nodes and the second-order neighbor nodes as well as the relation features, in order to ensure the accuracy of subsequent representation information acquisition.

[0133] S405 uses the first target linear transformation matrix and the second target linear transformation matrix of the spatial relationship perception layer to perform linear transformation on the spatial distance representation vector and the second feature vector, thereby determining the spatial weight vector of the node under the relationship path.

[0134] It is understandable that the sample representation vector of the sample space distance is a vector representing the characteristics of the sample space. In some implementations, the second objective linear transformation matrix of the spatial relationship perception layer is also obtained during the training process of the objective space perception model.

[0135] Alternatively, the spatial weight vector can be obtained as follows:

[0136]

[0137] in, The vector represents the spatial weight vector, where r1 and r2 represent the relationships in the relational path, namely the relationship between a node and its first-order neighbor nodes, and the relationship between a first-order neighbor node and its second-order neighbor nodes, respectively. Represents the linear transformation matrix of the target space, used for... The fusion and transformation can be performed again, which can be obtained during the training process of the target space perception model; d represents the linear transformation matrix of the second objective; i,j A vector representing spatial distance; h represents the first objective linear transformation matrix; k This represents the second feature vector of a second-order neighbor node; This represents the concatenation symbol, used to concatenate vectors.

[0138] In other words, the second feature vector and the representation vector are multiplied by the first target linear transformation matrix and the second target linear transformation matrix, respectively. The multiplied vectors are then concatenated to obtain the spatial weight vector, which fully reflects the spatial characteristics of the node under the relational path. Based on the spatial weight vector, the representation information is obtained more accurately.

[0139] S406 fuses the spatial weight vector and the relational weight vector, and then performs activation function mapping after multiplying the fused vector with the first target model parameters of the spatial relation perception layer to obtain the weight coefficients of the second-order neighbor nodes.

[0140] Alternatively, the weighting coefficients can be obtained using the following methods:

[0141]

[0142] Where Φ(i,j,k) represents the weight coefficient, and i,j,k are the node, the first-order neighbor node of the node, and the second-order neighbor node of the node, respectively. This represents the parameters of the first target model, which can be obtained during the training of the target spatial relationship perception model. Represents the relation weight vector; represents the spatial weight vector; sigmoid represents the activation function.

[0143] Alternatively, the activation function can also be the tanh function or the ReLU function, which is not limited here.

[0144] This weighting coefficient integrates spatial and relational information within the second-order neighborhood, thus more accurately measuring the importance and influence of second-order neighbor nodes.

[0145] S407, based on the first objective linear transformation matrix and weight coefficients, fuse the first eigenvector and the second eigenvector to obtain the third POI representation information of the node under the relation path.

[0146] Optionally, the first feature vector can be linearly transformed based on the first target linear transformation matrix to obtain the first transformation result; the second feature vector and the weight coefficients can be weighted to obtain the weighted second feature vector, and the weighted second feature vector can be linearly transformed based on the first target linear transformation matrix to obtain the second transformation result; the first transformation result and the second transformation result can be added together to obtain the third POI representation information of the node under the relation path.

[0147] For example, the acquisition of the third POI characterization information can be achieved as follows:

[0148]

[0149] in, Indicates the third POI characterization information; h j This represents the first eigenvector; Represents the linear transformation matrix of the first objective; The first transformation result is represented by Φ(i,j,k); the weighting coefficients are represented by h. k This represents the second eigenvector; This indicates the result of the second transformation.

[0150] Since the weight coefficients contain spatial and relational information, the first and second eigenvectors are fused using the weight coefficients and the first target linear transformation matrix. This effectively preserves the spatial and relational information, resulting in richer and more comprehensive third POI feature information, which allows for a more thorough mining of the representational information of the current node under the relational path.

[0151] S408, the first fusion layer in the target space relationship perception model performs a fusion operation on all the third POI representation information of the node under the same relationship pattern to obtain the second POI representation information of the node under the same relationship pattern.

[0152] Optionally, since the third POI representation information of a node under each relation path is also in vector form, the first fusion layer in the target space relation perception model performs a fusion operation on all the third POI representation information of the node under the same relation pattern. This can be a weighted summation of all the third POI representation information or an average calculation of all the third POI representation information. In other words, the third POI representation information of the node in all relation paths under the same relation pattern is aggregated to obtain the second POI representation information of the node under the same relation pattern. This second POI representation information also contains relation information and spatial information, thus fully mining the representation information of the node under the relation pattern.

[0153] S409, the second fusion layer in the target space relationship perception model performs a fusion operation on the second POI representation information of the node in each relationship mode to obtain the first POI representation information of the node in all relationship modes.

[0154] Furthermore, after determining the second POI representation information of a node under the same relational pattern based on the first fusion layer in the target spatial relation perception model, the second fusion layer is used to perform another fusion operation on the second POI representation information of the node under each relational pattern to obtain the first POI representation information of the node under all relational patterns. That is, a weighted summation operation or an average operation is performed on all the second POI representation information to obtain the first POI representation information of the node under all relational patterns.

[0155] like Figure 4aAs shown, in this embodiment of the disclosure, spatial distance is calculated using POI spatial location information, and second-order neighborhood relationships of nodes are mined using the POI relationship graph. The mined second-order neighborhood relationships and spatial distances are input into the target spatial relationship perception model to obtain POI representations, which are the first POI representation information. The first POI representation information retains both the complex relationships between nodes and the spatial characteristic information between nodes, making the information mining of POI nodes more thorough and the first POI representation information more accurate.

[0156] In this embodiment, a pre-trained target spatial relationship perception model is acquired, and the first POI representation information is obtained using this model. This results in faster processing speed and more stable and accurate results. The first and second feature vectors are processed using the first target linear transformation matrix in the target spatial relationship perception model to obtain the relationship weight vector under the relationship path, preserving the relationship information between nodes. The representation vector and the second feature vector are processed using the second and first target linear transformation matrices in the target spatial relationship perception model to determine the spatial weight vector, thus preserving the spatial characteristic information between nodes more completely. Furthermore, weight coefficients are determined using the first target model parameters, the relationship weight vector, and the spatial weight vector, reasonably fusing the relationship information and spatial information. The weight coefficients are used to obtain the third POI representation information, and this third POI representation information is then fused to obtain the final first POI representation information. The mining of the first POI representation information preserves the complex relationships between nodes and also retains the spatial characteristic relationships between nodes, resulting in more comprehensive and accurate first POI representation information.

[0157] Based on the above embodiments, the training process of the spatial relationship perception model of POIs in the method for obtaining representation information of points of interest (POIs) will be described. Figure 5 This is a schematic diagram illustrating a training method for a spatial relationship-aware model of Points of Interest (POIs) provided in an embodiment of this disclosure. Figure 5 As shown, the method includes, but is not limited to, the following steps:

[0158] S501, obtain the sample POI relationship graph, and based on the sample POI relationship graph, obtain the sample node relationship set of each sample node on the sample POI relationship graph.

[0159] The sample POI relationship graph includes sample nodes representing POIs and edges representing the relationships between sample nodes. The sample node relationship set includes the relationship paths from a sample node to all second-order neighbor nodes via any first-order neighbor node.

[0160] Optionally, a sample POI relationship graph can be determined based on an existing map, which contains different sample nodes and the relationships between sample nodes.

[0161] In the sample POI relationship graph, each sample node represents a different sample POI, and there are certain relationships between connected sample nodes. Based on the connections between each sample node and other sample nodes, the set of sample node relationships corresponding to each sample node is determined.

[0162] Optionally, the sample relation set includes relational paths from a sample node to all second-order neighbor nodes via any first-order neighbor node.

[0163] S502, obtain the sample space distance matrix between the sample node and its corresponding second-order neighbor node.

[0164] Optionally, the sample space distance between the sample node and its corresponding second-order neighbor node is obtained based on the coordinate information of the sample node and the coordinate information of its second-order neighbor node. The sample space distance can be Euclidean distance, Manhattan distance, or Chebyshev distance, etc.

[0165] In this embodiment, Euclidean distance is used as the sample space distance. The corresponding sample space distance is obtained based on the coordinate information of the sample node and its corresponding second-order neighbor node, and the corresponding sample space distance matrix is ​​obtained based on all the sample space distances corresponding to the sample node.

[0166] S503 uses the first linear transformation matrix of the spatial relationship perception layer to perform a linear transformation on the first sample feature vector of the first-order neighbor node and the second sample feature vector of the second-order neighbor node of the sample node, thereby determining the sample relationship weight vector of the sample node under the relationship path.

[0167] It is understandable that the first sample feature vector of a first-order neighbor node is a vector representing the features of that first-order neighbor node, and the second sample feature vector of a second-order neighbor node is a vector representing the features of that second-order neighbor node; the first sample feature vector and the second sample feature vector can be determined based on the category of the corresponding sample node.

[0168] In some implementations, the first linear transformation matrix can be a randomly generated initialization matrix, and the first linear transformation matrix is ​​optimized and adjusted during the continuous training of the spatial relationship-aware model.

[0169] Alternatively, the sample relationship weight vector can be obtained by:

[0170]

[0171] in, The sample relationship weight vector is represented by r1 and r2, which represent the relationships in the relationship path, namely the relationship between a node and its first-order neighbor node, and the relationship between a first-order neighbor node and its second-order neighbor node, respectively. Represents a linear transformation matrix of relations, used for... The fusion and transformation can be performed again to initialize the randomly generated matrix, and it can be optimized and adjusted during the continuous training of the spatial relationship perception model. h represents the first linear transformation matrix; j ′ represents the first sample feature vector of the first-order neighbor node; h k ′ represents the second sample feature vector of the second-order neighbor node; This represents the concatenation symbol, used to concatenate vectors.

[0172] S504 uses the first linear transformation matrix and the second linear transformation matrix of the spatial relationship perception layer to perform a linear transformation on the sample representation vector of the sample spatial distance and the second sample feature vector, thereby determining the sample spatial weight vector of the sample node under the relationship path.

[0173] It is understandable that the sample representation vector of the sample space distance is a vector representing the characteristics of the sample space. In some implementations, the second linear transformation matrix can be a randomly generated initialization matrix, and the second linear transformation matrix is ​​optimized and adjusted during the continuous training of the spatial relationship-aware model.

[0174] Alternatively, the sample space weight vector can be obtained by:

[0175]

[0176] in, R1 and R2 represent the weight vector of the sample space, respectively, which are the relationships in the relation path, that is, the relationship between a node and its first-order neighbor node, and the relationship between a first-order neighbor node and its second-order neighbor node. Represents a spatial linear transformation matrix, used for... The fusion and transformation can be performed again to initialize the randomly generated matrix, and it can be optimized and adjusted during the continuous training of the spatial relationship perception model. d represents the second linear transformation matrix; i,j ′ represents the sample representation vector of the sample spatial distance; h represents the first linear transformation matrix; k ′ represents the second feature vector of the second-order neighbor node; This represents the concatenation symbol, used to concatenate vectors.

[0177] S505 fuses the sample space weight vector and the sample relationship weight vector, and then performs activation function mapping after multiplying the fused sample vector with the first model parameters of the spatial relationship perception layer to obtain the sample weight coefficients of the second-order neighbor nodes.

[0178] In some implementations, fusing the sample space weight vector and the sample relation weight vector can be achieved by adding the sample space weight vector and the sample relation weight vector together to obtain sample weight coefficients that contain both spatial and relational information.

[0179] In some implementations, the activation function can be the sigmoid function, the tanh function, or the ReLU function.

[0180] Alternatively, the sample weight coefficients can be obtained by:

[0181]

[0182] Where Φ(i,j,k)′ represents the sample weight coefficient, and i,j,k are the node, the first-order neighbor node of the node, and the second-order neighbor node of the node, respectively. This represents the first model parameter, which can be obtained during the training of the target spatial relationship perception model. Represents the sample relationship weight vector; represents the sample space weight vector; sigmoid represents the activation function.

[0183] S506, based on the first linear transformation matrix and sample weight coefficients, fuse the first feature vector and the second feature vector to obtain the third sample POI representation information of the sample node under the relation path.

[0184] Optionally, the first sample feature vector can be linearly transformed based on the first linear transformation matrix to obtain the first transformation result; the second sample feature vector and the sample weight coefficient can be weighted to obtain the weighted second sample feature vector, and the weighted second sample feature vector can be linearly transformed based on the first linear transformation matrix to obtain the second transformation result; the first transformation result and the second transformation result can be added together to obtain the third sample POI representation information of the sample node under the relation path.

[0185] For example, the acquisition of the third POI characterization information can be achieved as follows:

[0186]

[0187] in, Indicates the POI characterization information of the third sample; h j ′ represents the first eigenvector; Denotes the first linear transformation matrix; The first transformation result is represented by Φ(i,j,k)′; the sample weight coefficients are represented by h. k ′ represents the second eigenvector; This indicates the result of the second transformation.

[0188] S507, the first fusion layer in the spatial relationship perception model performs a fusion operation on all the third sample POI representation information of the sample node under the same relationship pattern to obtain the second sample POI representation information of the sample node under the same relationship pattern.

[0189] It is understandable that the third sample POI representation information of the sample node under each relation path is actually a vector. Therefore, the first fusion layer in the spatial relation perception model is used to perform a fusion operation on all the third sample POI representation information of the sample node under the same relation pattern.

[0190] Optionally, the fusion operation of the representation information of all third sample POIs of the sample node under the same relational pattern can be either a weighted summation operation of all third sample POI representation information or an average operation of all third sample POI representation information.

[0191] By fusing the representation information of all third sample POIs of the sample node under the same relational pattern, the representation information of the second sample POI of the sample node under the same relational pattern is obtained.

[0192] S508, the second fusion layer in the target space relationship perception model performs a fusion operation on the second sample POI representation information of the sample node in each relationship mode to obtain the first sample POI representation information of the sample node in all relationship modes.

[0193] It is understandable that the second sample POI representation information of the sample node in each relation pattern is actually also a vector. Therefore, when determining the first sample POI representation information of the sample node, the second fusion layer in the spatial relation perception model can be used to perform a fusion operation on the second sample POI representation information of the sample node in all relation patterns.

[0194] Optionally, the fusion operation of the second sample POI representation information of the sample node under all relational patterns can be either a weighted summation operation of all the second sample POI representation information or an average operation of all the second sample POI representation information.

[0195] By fusing the second sample POI representation information of the sample node under all relational schemas, the first sample POI representation information of the sample node under all relational schemas is obtained.

[0196] S509: Based on the POI representation information of the first sample, the downstream task prediction data of the sample node is obtained. Based on the downstream task prediction data and the downstream task annotation data of the sample node, the spatial relationship perception model is corrected until the training ends and the target spatial relationship perception model is obtained.

[0197] Alternatively, the downstream task is the actual task that you want to solve. For example, in a location recommendation scenario, the downstream task is to obtain the final recommended location.

[0198] In some implementations, after determining the first sample POI representation information of a sample node, the downstream task prediction data of the sample node can be determined. For example, in a location recommendation scenario, the predicted recommendation value of a sample node can be determined based on the first sample POI representation information of that sample node. The predicted recommendation value is the downstream task prediction data.

[0199] Optionally, based on downstream task prediction data and downstream task annotation data of sample nodes, a loss function for the spatial relationship perception model is determined; based on the loss function, the first linear transformation matrix, the second linear transformation matrix, and the first model parameters in the spatial relationship perception model are corrected, and the next sample is returned to continue training the corrected spatial relationship perception model until the training ends and the target spatial relationship perception model is obtained. The target spatial relationship perception model includes the first target linear transformation matrix, the second target linear transformation matrix, and the first target model parameters.

[0200] Optionally, the loss function of the spatial relationship-aware model can be the absolute difference between the downstream task prediction data and the downstream task annotation data of the sample node. It is understood that the smaller the loss function value, the more likely the spatial relationship-aware model is to converge; that is, the absolute difference reflects the difference between the downstream task prediction data and the actual downstream task annotation data of the sample node. The larger the absolute difference, the greater the difference between the downstream task prediction data and the actual downstream task annotation data of the sample node, and the worse the spatial relationship-aware model's performance in acquiring the representation information of the first sample POI.

[0201] For example, assuming the downstream task is location recommendation, the predicted data for the downstream task is the score of the sample POI node. The score of the sample POI node can be obtained by similarity matching between the first sample POI representation information output by the spatial relationship awareness model and the task representation information. The loss function between the score of the sample POI node and the label value of the sample POI node is calculated. The label value is the labeled data of the downstream task. When the loss function converges, the training of the spatial awareness model is completed.

[0202] For example, suppose the downstream task is to perform competition analysis on sample POI nodes. The downstream task prediction data is the first sample POI representation information of the sample POI nodes. The downstream task annotation data can be the label value indicating whether there is a competitive relationship between sample POI nodes. The label value can be 0 or 1, where 0 represents no competition and 1 represents competition. The loss function is calculated using the first sample POI representation information and the label value of the sample POI nodes. When the loss function converges, the training of the spatial perception model is completed.

[0203] It is understandable that the first linear transformation matrix, the second linear transformation matrix, and the first model parameters in the spatial relationship perception model are corrected. Based on the corrected first linear transformation matrix, the second linear transformation matrix, and the first model parameters, the new first sample POI representation information of the sample node can be determined. Based on the new first sample POI representation information, new downstream task prediction data is obtained, and a new loss function is obtained based on the new downstream task prediction data.

[0204] It should be noted that if the process of obtaining the sample relation weight vector and the sample space weight vector includes a relation linear transformation matrix and a space linear transformation matrix, then the relation linear transformation matrix and the space linear transformation matrix also need to be corrected, and the sample relation weight vector and the sample space weight vector are obtained based on the corrected relation linear transformation matrix and the space linear transformation matrix for subsequent calculations.

[0205] By continuously modifying the first linear transformation matrix, the second linear transformation matrix, and the first model parameters in the spatial relationship perception model, the result of the loss function is continuously changed until the loss function converges. At this point, the modification of the spatial relationship perception model ends, which is also the end of the training of the spatial relationship perception model, resulting in the target spatial relationship perception model. The target spatial relationship perception model at this point includes the first target linear transformation matrix, the second target linear transformation matrix, and the first target model parameters.

[0206] It should be noted that if the acquisition of the sample relation weight vector and the sample space weight vector includes the relation linear transformation matrix and the space linear transformation matrix, then the target space perception model also includes the target relation linear transformation matrix and the target space linear transformation matrix.

[0207] In this embodiment, a set of sample node relationships is obtained by acquiring a sample POI relationship graph; based on the relationship paths in the sample node relationship graph and the spatial distance between sample nodes, first sample POI representation information is obtained, and downstream task prediction data is obtained based on the first sample POI representation information; a loss function is obtained based on the downstream task prediction data and the actual downstream task annotation data; the first linear transformation matrix, the second linear transformation matrix, and the first model parameters in the spatial relationship perception model are continuously corrected through the loss function so that the loss function of the spatial relationship perception model converges; the spatial relationship perception model is trained using the relationship information and spatial information in the actual sample POI relationship graph, so that the spatial relationship perception model can better learn the complex relationships and spatial characteristics between sample nodes, and the effect of the trained target spatial relationship perception model is better.

[0208] Figure 6 This is a schematic diagram of a method for obtaining representational information of Points of Interest (POIs) provided in an embodiment of this disclosure. Figure 6 As shown, the method includes, but is not limited to, the following steps:

[0209] S601, obtain the sample POI relationship graph, and based on the sample POI relationship graph, obtain the sample node relationship set of each sample node on the sample POI relationship graph.

[0210] The sample POI relationship graph includes sample nodes representing POIs and edges representing the relationships between sample nodes. The sample node relationship set includes the relationship paths from a sample node to all second-order neighbor nodes via any first-order neighbor node.

[0211] In this embodiment of the disclosure, step S601 can be implemented in any of the embodiments of this disclosure, and no limitation is made here, nor will it be described in detail.

[0212] S602, obtain the sample space distance matrix between the sample node and its corresponding second-order neighbor node.

[0213] In this embodiment of the disclosure, step S602 can be implemented in any of the embodiments of this disclosure, and no limitation is made here, nor will it be described in detail.

[0214] S603 uses the first linear transformation matrix of the spatial relationship perception layer to perform a linear transformation on the first sample feature vector of the first-order neighbor node and the second sample feature vector of the second-order neighbor node of the sample node, thereby determining the sample relationship weight vector of the sample node under the relationship path.

[0215] In this embodiment of the disclosure, step S603 can be implemented in any of the embodiments of this disclosure, and no limitation is made here, nor will it be described in detail.

[0216] S604 uses the first linear transformation matrix and the second linear transformation matrix of the spatial relationship perception layer to perform a linear transformation on the sample representation vector of the sample spatial distance and the second sample feature vector, thereby determining the sample spatial weight vector of the sample node under the relationship path.

[0217] In this embodiment of the disclosure, step S604 can be implemented in any of the embodiments of this disclosure, and no limitation is made here, nor will it be described in detail.

[0218] S605 fuses the sample space weight vector and the sample relationship weight vector, and then performs activation function mapping after multiplying the fused sample vector with the first model parameters of the spatial relationship perception layer to obtain the sample weight coefficients of the second-order neighbor nodes.

[0219] In this embodiment of the disclosure, step S605 can be implemented in any of the embodiments of this disclosure, and no limitation is made here, nor will it be described in detail.

[0220] S606, based on the first linear transformation matrix and sample weight coefficients, fuses the first feature vector and the second feature vector to obtain the third sample POI representation information of the sample node under the relation path.

[0221] In this embodiment of the disclosure, step S606 can be implemented in any of the various embodiments of the disclosure, and no limitation is made here, nor will it be described in detail.

[0222] S607, the first fusion layer in the spatial relationship perception model performs a fusion operation on all the third sample POI representation information of the sample node under the same relationship pattern to obtain the second sample POI representation information of the sample node under the same relationship pattern.

[0223] In this embodiment of the disclosure, step S607 can be implemented in any of the embodiments of this disclosure, and no limitation is made here, nor will it be described in detail.

[0224] S608, the second fusion layer in the target space relationship perception model performs a fusion operation on the second sample POI representation information of the sample node in each relationship mode to obtain the first sample POI representation information of the sample node in all relationship modes.

[0225] In this embodiment of the disclosure, step S608 can be implemented in any of the embodiments of this disclosure, and no limitation is made here, nor will it be described in detail.

[0226] S609: Based on the POI representation information of the first sample, the downstream task prediction data of the sample node is obtained. Based on the downstream task prediction data and the downstream task annotation data of the sample node, the spatial relationship perception model is corrected until the training ends and the target spatial relationship perception model is obtained.

[0227] In this embodiment of the disclosure, step S609 can be implemented in any of the embodiments of this disclosure, and no limitation is made here, nor will it be described in detail.

[0228] S610, obtain the POI relationship graph, and based on the POI relationship graph, obtain the set of node relationships for each node in the POI relationship graph.

[0229] The POI relationship graph includes nodes representing POIs and edges representing relationships between nodes. The node relationship set includes the relationship paths from a node to all second-order neighbors via any first-order neighbor node.

[0230] In this embodiment of the disclosure, step S610 can be implemented in any of the various embodiments of the disclosure, and no limitation is made here, nor will it be described in detail.

[0231] S611, obtain the spatial distance matrix between the node and its corresponding second-order neighbor nodes.

[0232] In this embodiment of the disclosure, step S611 can be implemented in any of the various embodiments of the disclosure, and no limitation is made here, nor will it be described in detail.

[0233] S612 uses the first target linear transformation matrix of the spatial relationship perception layer to perform linear transformation on the first feature vector of the first-order neighbor node and the second feature vector of the second-order neighbor node of the node, and determines the relationship weight vector of the node under the relationship path.

[0234] In this embodiment of the disclosure, step S612 can be implemented in any of the embodiments of this disclosure, and no limitation is made here, nor will it be described in detail.

[0235] S613, using the first target linear transformation matrix and the second target linear transformation matrix of the spatial relationship perception layer, performs a linear transformation on the spatial distance representation vector and the second feature vector to determine the spatial weight vector of the node under the relationship path.

[0236] In this embodiment of the disclosure, step S613 can be implemented in any of the embodiments of this disclosure, and no limitation is made here, nor will it be described in detail.

[0237] S614 fuses the spatial weight vector and the relational weight vector, and then performs activation function mapping after multiplying the fused vector with the first target model parameters of the spatial relation perception layer to obtain the weight coefficients of the second-order neighbor nodes.

[0238] In this embodiment of the disclosure, step S614 can be implemented in any of the various embodiments of the disclosure, and no limitation is made here, nor will it be described in detail.

[0239] S615, based on the first objective linear transformation matrix and weight coefficients, fuse the first eigenvector and the second eigenvector to obtain the third POI representation information of the node under the relation path.

[0240] In this embodiment of the disclosure, step S615 can be implemented in any of the embodiments of this disclosure, and no limitation is made here, nor will it be described in detail.

[0241] S616, the first fusion layer in the target space relationship perception model performs a fusion operation on all the third POI representation information of the node under the same relationship pattern to obtain the second POI representation information of the node under the same relationship pattern.

[0242] In this embodiment of the disclosure, step S616 can be implemented in any of the various embodiments of the disclosure, and no limitation is made here, nor will it be described in detail.

[0243] S617, the second fusion layer in the target space relationship perception model performs a fusion operation on the second POI representation information of the node in each relationship mode to obtain the first POI representation information of the node in all relationship modes.

[0244] In this embodiment of the disclosure, step S617 can be implemented in any of the embodiments of this disclosure, and no limitation is made here, nor will it be described in detail.

[0245] In this embodiment, the first sample POI representation information is obtained by utilizing the relationship information and spatial characteristic information in the sample POI relationship graph. This information is then used to train the spatial relationship perception model, enabling the model to better learn the complex relationships and spatial characteristics between nodes. In practical applications, the node relationship set and spatial distance matrix in the POI relationship graph are obtained and input into the pre-trained target spatial relationship perception model to obtain more accurate first POI representation information. Furthermore, during the acquisition of the first POI representation information, it is obtained through different relationship paths. The third POI representation information under the same relationship pattern is fused to obtain the second POI representation information. The second POI representation information under each relationship pattern is then fused again to obtain the final first POI representation information. This approach comprehensively mines the second-order neighborhood information of nodes, retains relationship information and spatial characteristic information, and makes the first POI representation information of nodes more accurate across all relationship patterns.

[0246] Based on the above embodiments, the embodiments of this disclosure can be applied to POI recommendation scenarios, obtaining each POI node to be recommended; determining the first POI representation information of the POI node to be recommended using a target spatial relationship perception model based on the complex relationship information and spatial distance information between the POI node and its second-order neighborhood; inputting the first representation information of the POI node to be recommended into the recommendation model of the downstream task, and having the recommendation model score the POI node based on the first representation information of the POI node. The score of the POI node to be recommended can be considered as the score at which the POI node should be recommended. The higher the score of the POI node to be recommended, the more it should be recommended; for example, the scores of all the POI nodes to be recommended can be sorted from largest to smallest, and the top K POI nodes to be recommended after sorting can be used as recommended POIs.

[0247] Based on the above embodiments, the present disclosure embodiments can also be applied to the identification of POI node competition relationships. After obtaining the first POI representation information of each POI node, the first POI representation information of the POI node is input into the competition relationship identification model of the downstream task. The competition relationship identification model analyzes the first POI representation information of the node and outputs the relationship result. The relationship result can have different values. For example, if the relationship result is 1, it can be a competition relationship; if the relationship result is 0, it can be a complementary relationship; if the relationship result is 2, it can be a dependency relationship, etc.

[0248] Figure 7 This is a structural block diagram of a device for acquiring representational information of Points of Interest (POIs) provided in an embodiment of this disclosure. Figure 7 As shown, the device 700 for acquiring the representation information of the point of interest (POI) includes:

[0249] The first acquisition module 701 is used to acquire the POI relationship graph and, based on the POI relationship graph, acquire the node relationship set of each node on the POI relationship graph. The POI relationship graph includes nodes representing POIs and edges between nodes, and the node relationship set includes the relationship paths of a node reaching all second-order neighbor nodes through any first-order neighbor node.

[0250] The second acquisition module 702 is used to acquire the spatial distance matrix between a node and its corresponding second-order neighbor nodes.

[0251] The information acquisition module 703 is used to determine the first POI representation information of a node based on the node relationship set and spatial distance matrix.

[0252] In some implementations, the information acquisition module 703 includes:

[0253] From the set of node relations, determine the set of relations of a node under different relation schemas. The set of relations includes at least one relation path of a node under the same relation schema.

[0254] For each relation schema, the second POI representation information of the node in the relation schema is determined based on the relation path and spatial distance matrix in the relation subset;

[0255] The second POI representation information of the node in each relation schema is fused to obtain the first POI representation information of the node in all relation schemas.

[0256] In some implementations, the information acquisition module 703 includes:

[0257] For each relation path in the relation subset corresponding to the same relation schema, the spatial distance between the node in the relation path and its second-order neighbor node is determined based on the spatial distance matrix.

[0258] Based on spatial distance and the first-order and second-order neighbor nodes of a node in the relational path, determine the third POI representation information of the node under the relational path;

[0259] The third POI representation information of a node under the same relation schema is fused to obtain the second POI representation information of the node under the same relation schema.

[0260] In some implementations, the information acquisition module 703 includes:

[0261] Based on first-order neighbor nodes and second-order neighbor nodes, determine the relation weight vector of a node under the relation path;

[0262] Based on spatial distance and second-order neighbor nodes, determine the spatial weight vector of a node under the relation path;

[0263] Based on first-order neighbor nodes and second-order neighbor nodes, determine the relation weight vector of a node under the relation path;

[0264] The spatial weight vector and the relational weight vector are fused to obtain the weight coefficients of the second-order neighbor nodes;

[0265] Based on first-order neighbor nodes, second-order neighbor nodes, and weight coefficients, the third POI representation information of the node under the relation path is obtained.

[0266] In some implementations, the information acquisition module 703 includes:

[0267] Determine the first feature vector of the first-order neighbor node and the second feature vector of the second-order neighbor node of the node in the relation path;

[0268] Perform a linear transformation on the first and second eigenvectors to determine the relation weight vectors corresponding to the nodes.

[0269] In some implementations, the information acquisition module 703 includes:

[0270] Determine the spatial distance representation vector and the second feature vector of the second-order neighbor nodes;

[0271] A linear transformation is performed on the spatial distance representation vector and the second feature vector to determine the spatial weight vector corresponding to the node.

[0272] In some implementations, the information acquisition module 703 includes:

[0273] Obtain a pre-trained target spatial relationship perception model;

[0274] Each relation path in the relation subset corresponding to the same relation schema, and the spatial distance between the node in the relation path and the second-order neighbor node, are input into the target spatial relation perception model. The spatial relation perception layer in the target spatial relation perception model outputs the third POI representation information of the node under the relation path.

[0275] The first fusion layer in the target space relationship perception model performs a fusion operation on all the third POI representation information of the node under the same relationship pattern to obtain the second POI representation information of the node under the same relationship pattern.

[0276] The second fusion layer in the target space relationship perception model performs a fusion operation on the second POI representation information of the node in each relationship mode to obtain the first POI representation information of the node in all relationship modes.

[0277] In some implementations, the information acquisition module 703 includes:

[0278] The first target linear transformation matrix of the spatial relationship perception layer is used to linearly transform the first feature vector of the first-order neighbor node and the second feature vector of the second-order neighbor node of the node to determine the relation weight vector of the node under the relation path.

[0279] The spatial distance representation vector and the second feature vector are linearly transformed by the first target linear transformation matrix and the second target linear transformation matrix of the spatial relationship perception layer to determine the spatial weight vector of the node under the relationship path.

[0280] The spatial weight vector and relational weight vector are fused, and the fused vector is multiplied with the first target model parameters of the spatial relation perception layer and then mapped by the activation function to obtain the weight coefficients of the second-order neighbor nodes.

[0281] Based on the first objective linear transformation matrix and weight coefficients, the first eigenvector and the second eigenvector are fused to obtain the third POI representation information of the node under the relation path.

[0282] In some implementations, the information acquisition module 703 includes:

[0283] A linear transformation is performed on the first eigenvector based on the first target linear transformation matrix to obtain the first transformation result;

[0284] The second eigenvector and the weight coefficients are weighted to obtain the weighted second eigenvector. The weighted second eigenvector is then linearly transformed based on the first target linear transformation matrix to obtain the second transformation result.

[0285] The first and second transformation results are added together to obtain the third POI representation information of the node under the relation path.

[0286] In some implementations, the first acquisition module 701 includes:

[0287] For any candidate second-order neighbor node vk of node vi, based on the POI relationship graph, obtain the number of relationship paths from node vi to node vk through any first-order neighbor node vj.

[0288] From all the candidate second-order neighbor nodes of node vi, determine one or more candidate second-order neighbor nodes whose number of paths is greater than a set number, and retain them as the target second-order neighbor nodes of the node.

[0289] The relationship paths between a node and its target second-order neighbor nodes are mined to obtain the set of node relationships.

[0290] In some implementations, the second acquisition module 702 includes:

[0291] For any second-order neighbor node of a node, obtain the first spatial coordinate information of the node and the second spatial coordinate information of the second-order neighbor node;

[0292] Euclidean distance calculation is performed based on the first spatial coordinate information and the second spatial coordinate information to obtain the spatial distance between the node and its second-order neighbor nodes.

[0293] Based on the spatial distance between a node and each of its second-order neighbor nodes, a spatial distance matrix is ​​obtained.

[0294] In some implementations, after the information acquisition module 703, the following are included:

[0295] Based on the first POI representation information of the node, the node is processed by downstream tasks to obtain the processing results of the downstream tasks.

[0296] In this embodiment, a POI relationship graph and a set of node relationships in the POI relationship graph are obtained. The set of node relationships includes all complex relationships between nodes. A spatial distance matrix is ​​determined based on the spatial location of the nodes. The spatial distance matrix reflects the spatial characteristic information between nodes. Based on the set of node relationships and the fixed spatial distance matrix, the first POI representation information is determined. This fully explores the relationship information and spatial information within the second-order neighborhood of the nodes, and the obtained first POI representation information is more accurate and representative.

[0297] Figure 8 This is a structural block diagram of a training device for a spatial relationship-aware model of points of interest (POIs) provided in an embodiment of this disclosure. Figure 8 As shown, the training device 800 for the spatial relationship perception model of the Point of Interest (POI) includes:

[0298] The first sample acquisition module 801 is used to acquire the sample POI relationship graph and, based on the sample POI relationship graph, acquire the sample node relationship set of each sample node on the sample POI relationship graph. The sample POI relationship graph includes sample nodes representing POIs and edges of relationships between sample nodes. The sample node relationship set includes the relationship paths of sample nodes reaching all second-order neighbor nodes through any first-order neighbor node.

[0299] The second sample acquisition module 802 is used to acquire the sample space distance matrix between the sample node and its corresponding second-order neighbor node.

[0300] The sample information acquisition module 803 is used to train the initial spatial relationship perception model based on the sample node relationship set and sample spatial distance matrix of the sample nodes, and to determine the first sample POI representation information of the sample nodes.

[0301] The correction module 804 is used to obtain the downstream task prediction data of the sample node based on the first sample POI representation information, and to correct the spatial relationship perception model based on the downstream task prediction data and the downstream task annotation data of the sample node until the training ends and the target spatial relationship perception model is obtained.

[0302] In some implementations, the sample information acquisition module 803 includes:

[0303] Each relation path in the relation subset corresponding to the same relation pattern, and the sample spatial distance between the sample node and the second-order neighbor node in the relation path, are input into the spatial relation perception model. The spatial relation perception layer in the spatial relation perception model outputs the third sample POI representation information of the sample node under the relation path.

[0304] The first fusion layer in the spatial relationship perception model performs a fusion operation on all third sample POI representation information of sample nodes under the same relationship pattern to obtain the second sample POI representation information of sample nodes under the same relationship pattern.

[0305] The second fusion layer in the target space relationship perception model performs a fusion operation on the second sample POI representation information of the sample node in each relationship pattern to obtain the first sample POI representation information of the sample node in all relationship patterns.

[0306] In some implementations, the sample information acquisition module 803 includes:

[0307] The first linear transformation matrix of the spatial relationship perception layer is used to linearly transform the first sample feature vector of the first-order neighbor node and the second sample feature vector of the second-order neighbor node of the sample node to determine the sample relationship weight vector of the sample node under the relationship path.

[0308] The first linear transformation matrix and the second linear transformation matrix of the spatial relationship perception layer are used to linearly transform the sample representation vector and the second sample feature vector of the sample spatial distance to determine the sample spatial weight vector of the sample node under the relationship path.

[0309] The sample space weight vector and the sample relationship weight vector are fused, and the sample fusion vector is multiplied with the first model parameters of the spatial relationship perception layer and then mapped by the activation function to obtain the sample weight coefficients of the second-order neighbor nodes.

[0310] Based on the first linear transformation matrix and sample weight coefficients, the first feature vector and the second feature vector are fused to obtain the third sample POI representation information of the sample node under the relation path.

[0311] In some implementations, the sample information acquisition module 803 includes:

[0312] The first sample feature vector is linearly transformed based on the first linear transformation matrix to obtain the first transformation result;

[0313] The second sample feature vector and sample weight coefficients are weighted to obtain the weighted second sample feature vector. The weighted second sample feature vector is then linearly transformed based on the first linear transformation matrix to obtain the second transformation result.

[0314] The first and second transformation results are added together to obtain the third sample POI representation information of the sample node under the relation path.

[0315] In some implementations, module 804 is corrected, including:

[0316] Based on downstream task prediction data and downstream task annotation data of sample nodes, the loss function of the spatial relationship perception model is determined.

[0317] The first linear transformation matrix, the second linear transformation matrix, and the first model parameters in the spatial relationship perception model are corrected based on the loss function. The next sample is then returned to continue training the corrected spatial relationship perception model until the training ends and the target spatial relationship perception model is obtained. The target spatial relationship perception model includes the first target linear transformation matrix, the second target linear transformation matrix, and the first target model parameters.

[0318] In this embodiment, a set of sample node relationships is obtained by acquiring a sample POI relationship graph; based on the relationship paths in the sample node relationship graph and the spatial distance between sample nodes, first sample POI representation information is obtained, and downstream task prediction data is obtained based on the first sample POI representation information; a loss function is obtained based on the downstream task prediction data and the actual downstream task annotation data; the first linear transformation matrix, the second linear transformation matrix, and the first model parameters in the spatial relationship perception model are continuously corrected through the loss function so that the loss function of the spatial relationship perception model converges; the spatial relationship perception model is trained using the relationship information and spatial information in the actual sample POI relationship graph, so that the spatial relationship perception model can better learn the complex relationships and spatial characteristics between sample nodes, and the effect of the trained target spatial relationship perception model is better.

[0319] The acquisition, storage, and application of user personal information involved in the technical solution disclosed herein comply with the provisions of relevant laws and regulations and do not violate public order and good morals.

[0320] According to embodiments of this disclosure, this disclosure also provides an electronic device, a readable storage medium, and a computer program product.

[0321] Figure 9A schematic block diagram of an example electronic device 900 that can be used to implement embodiments of the present disclosure is shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device may also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the present disclosure described and / or claimed herein.

[0322] like Figure 9 As shown, device 900 includes a computing unit 901, which can perform various appropriate actions and processes based on a computer program stored in read-only memory (ROM) 902 or a computer program loaded into random access memory (RAM) 903 from storage unit 908. RAM 903 may also store various programs and data required for the operation of device 900. The computing unit 901, ROM 902, and RAM 903 are interconnected via bus 904. Input / output (I / O) interface 905 is also connected to bus 904.

[0323] Multiple components in device 900 are connected to I / O interface 905, including: input unit 906, such as keyboard, mouse, etc.; output unit 907, such as various types of monitors, speakers, etc.; storage unit 908, such as disk, optical disk, etc.; and communication unit 909, such as network card, modem, wireless transceiver, etc. Communication unit 909 allows device 900 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0324] The computing unit 901 can be various general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the computing unit 901 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. The computing unit 901 performs the various methods and processes described above, such as methods for acquiring representation information of points of interest (POIs) and methods for training spatial relationship-aware models of POIs. For example, in some embodiments, the methods for acquiring representation information of POIs or training spatial relationship-aware models of POIs can be implemented as computer software programs tangibly contained in a machine-readable medium, such as storage unit 908. In some embodiments, part or all of the computer program can be loaded and / or installed on device 900 via ROM 902 and / or communication unit 909. When the computer program is loaded into RAM 903 and executed by computing unit 901, one or more steps of the method for acquiring representation information of points of interest (POIs) or the method for training a spatial relationship-aware model of POIs described above can be performed. Alternatively, in other embodiments, computing unit 901 can be configured to perform the method for acquiring representation information of POIs by any other suitable means (e.g., by means of firmware).

[0325] Various embodiments of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), systems-on-a-chip (SoCs), payload-programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0326] The program code used to implement the methods of this disclosure may be written in any combination of one or more programming languages. This program code may be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing apparatus, such that when executed by the processor or controller, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code may be executed entirely on a machine, partially on a machine, as a standalone software package partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0327] In the context of this disclosure, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium can be, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0328] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device for displaying information to the user (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor); and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the computer. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0329] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as a data server), or computing systems that include middleware components (e.g., an application server), or computing systems that include frontend components (e.g., a user computer with a graphical user interface or web browser through which a user can interact with embodiments of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., a communication network). Examples of communication networks include local area networks (LANs), wide area networks (WANs), and the Internet.

[0330] Computer systems can include clients and servers. Clients and servers are generally located far apart and typically interact via communication networks. Client-server relationships are created by computer programs running on the respective computers and having a client-server relationship with each other. Servers can be cloud servers, servers in distributed systems, or servers incorporating blockchain technology.

[0331] It should be understood that the various forms of processes shown above can be used to rearrange, add, or delete steps. For example, the steps described in this disclosure can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this disclosure can be achieved, and this is not limited herein.

[0332] The specific embodiments described above do not constitute a limitation on the scope of protection of this disclosure. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this disclosure should be included within the scope of protection of this disclosure.

Claims

1. A method for obtaining representation information of Points of Interest (POIs), wherein, The method includes: The POI relationship graph is determined based on the map, and the node relationship set of each node on the POI relationship graph is obtained based on the POI relationship graph. The POI relationship graph includes nodes representing POIs and edges between nodes, and the node relationship set includes the relationship paths of a node to all second-order neighbor nodes through any first-order neighbor node. Obtain the spatial distance matrix between the node and its corresponding second-order neighbor nodes, wherein the spatial distance matrix is ​​determined based on the spatial position of the node; Based on the node relationship set of the node and the spatial distance matrix, the first POI representation information of the node is determined; Wherein, after determining the first POI representation information of the node based on the node relationship set and the spatial distance matrix, the process includes: Based on the first POI representation information of the node, downstream task processing is performed on the node to obtain the processing result of the downstream task.

2. The method according to claim 1, wherein, The determination of the POI representation information of the node based on the node relationship set and the spatial distance matrix includes: From the set of node relationships, determine a subset of relationships for the node under different relationship patterns, wherein the subset of relationships includes at least one relationship path for the node under the same relationship pattern; For each of the relation patterns, based on the relation paths in the relation subset and the spatial distance matrix, the second POI representation information of the node under the relation pattern is determined; The second POI representation information of the node in each relation schema is fused to obtain the first POI representation information of the node in all relation schemas.

3. The method according to claim 2, wherein, The step of determining the second POI representation information of the node under the relation pattern based on the relation paths in the relation subset and the spatial distance matrix includes: For each relation path in the relation subset corresponding to the same relation schema, the spatial distance between the node in the relation path and its second-order neighbor node is determined based on the spatial distance matrix. Based on the spatial distance and the first-order and second-order neighbor nodes of the node in the relation path, determine the third POI representation information of the node under the relation path; The third POI representation information of the node under the same relation schema is fused to obtain the second POI representation information of the node under the same relation schema.

4. The method according to claim 3, wherein, The method of determining the third POI representation information of a node under the relation path based on the spatial distance and the first-order and second-order neighbor nodes of the node in the relation path includes: Based on the first-order neighbor node and the second-order neighbor node, determine the relation weight vector of the node under the relation path; Based on the spatial distance and the second-order neighbor nodes, determine the spatial weight vector of the node under the relation path; The spatial weight vector and the relational weight vector are fused to obtain the weight coefficients of the second-order neighbor nodes; Based on the first-order neighbor node, the second-order neighbor node, and the weight coefficient, the third POI representation information of the node under the relation path is obtained.

5. The method according to claim 4, wherein, The step of determining the relation weight vector of a node under the relation path based on the first-order neighbor node and the second-order neighbor node includes: Determine the first feature vector of the first-order neighbor node and the second feature vector of the second-order neighbor node of the node in the relation path; Perform a linear transformation on the first feature vector and the second feature vector to determine the relation weight vector corresponding to the node.

6. The method according to claim 4, wherein, The step of determining the spatial weight vector of a node under the relational path based on the spatial distance and the second-order neighbor nodes includes: Determine the representation vector of the spatial distance and the second feature vector of the second-order neighbor nodes; A linear transformation is performed on the spatial distance representation vector and the second feature vector to determine the spatial weight vector corresponding to the node.

7. The method according to claim 1, wherein, The determination of the POI representation information of the node based on the node relationship set and the spatial distance matrix includes: Obtain a pre-trained target spatial relationship perception model; Each relation path in the relation subset corresponding to the same relation pattern, and the spatial distance between the node in the relation path and the second-order neighbor node, are input into the target spatial relation perception model. The spatial relation perception layer in the target spatial relation perception model outputs the third POI representation information of the node under the relation path. The first fusion layer in the target spatial relationship perception model performs a fusion operation on all the third POI representation information of the node under the same relationship pattern to obtain the second POI representation information of the node under the same relationship pattern. The second fusion layer in the target spatial relationship perception model performs a fusion operation on the second POI representation information of the node in each relationship mode to obtain the first POI representation information of the node in all relationship modes.

8. The method according to claim 7, wherein, The step of inputting each relation path in the relation subset corresponding to the same relation pattern, and the spatial distance between nodes in the relation path and their second-order neighbor nodes, into the target spatial relation perception model, yields the third POI representation information of the node under the relation path, including: The first target linear transformation matrix of the spatial relationship perception layer is used to linearly transform the first feature vector of the first-order neighbor node and the second feature vector of the second-order neighbor node of the node to determine the relationship weight vector of the node under the relationship path. The spatial distance representation vector and the second feature vector are linearly transformed using the first target linear transformation matrix and the second target linear transformation matrix of the spatial relationship perception layer to determine the spatial weight vector of the node under the relationship path; The spatial weight vector and the relation weight vector are fused, and the fused vector is multiplied with the first target model parameters of the spatial relation perception layer and then mapped by an activation function to obtain the weight coefficients of the second-order neighbor nodes. Based on the first target linear transformation matrix and the weight coefficients, the first feature vector and the second feature vector are fused to obtain the third POI representation information of the node under the relation path.

9. The method according to claim 8, wherein, The process of fusing the first feature vector and the second feature vector based on the first target linear transformation matrix and the weight coefficients to obtain the third POI representation information of the node under the relation path includes: A first transformation result is obtained by performing a linear transformation on the first feature vector based on the first target linear transformation matrix. The second feature vector and the weight coefficient are weighted to obtain a weighted second feature vector, and a linear transformation is performed on the weighted second feature vector based on the first target linear transformation matrix to obtain a second transformation result; The first transformation result and the second transformation result are added together to obtain the third POI representation information of the node under the relationship path.

10. The method according to any one of claims 1-9, wherein, The step of obtaining the node relationship set for each node in the POI relationship graph based on the POI relationship graph includes: For the node v i any candidate second-order neighbor node v k Based on the POI relationship graph, obtain the nodes v i Passing through any first-order neighbor node v j Reaching the node v k The number of paths in the relational path; From the node v i Among all candidate second-order neighbor nodes, one or more candidate second-order neighbor nodes whose number of paths is greater than a set number are identified and retained as the target second-order neighbor nodes of the node. The relationship paths between the node and the target second-order neighbor node are mined to obtain the node relationship set of the node.

11. The method according to any one of claims 1-9, wherein, The step of obtaining the spatial distance matrix between the node and its corresponding second-order neighbor nodes includes: For any second-order neighbor node of the node, obtain the first spatial coordinate information of the node and the second spatial coordinate information of the second-order neighbor node; Euclidean distance calculation is performed based on the first spatial coordinate information and the second spatial coordinate information to obtain the spatial distance between the node and its second-order neighbor nodes. The spatial distance matrix is ​​obtained based on the spatial distance between the node and each of its second-order neighbor nodes.

12. A training method for a spatial relationship-aware model of Points of Interest (POIs), wherein, The method includes: The sample POI relationship graph is determined based on the map, and the sample node relationship set of each sample node on the sample POI relationship graph is obtained based on the sample POI relationship graph. The sample POI relationship graph includes sample nodes representing POIs and edges of relationships between sample nodes. The sample node relationship set includes relationship paths from a sample node to all second-order neighbor nodes via any first-order neighbor node. Obtain the sample spatial distance matrix between the sample node and its corresponding second-order neighbor node, wherein the spatial distance matrix is ​​determined based on the spatial position of the sample node; Based on the sample node relationship set and the sample spatial distance matrix, the initial spatial relationship perception model is trained to determine the first sample POI representation information of the sample node. The first POI representation information of the sample node is used to process the sample node for downstream tasks and obtain the processing result of the downstream task. Based on the first sample POI representation information, the downstream task prediction data of the sample node is obtained, and based on the downstream task prediction data and the downstream task annotation data of the sample node, the spatial relationship perception model is corrected until the training ends and the target spatial relationship perception model is obtained.

13. The method according to claim 12, wherein, The initial spatial relationship perception model is trained based on the sample node relationship set and the spatial distance matrix to determine the predicted POI representation information of the sample nodes, including: Each relation path in the relation subset corresponding to the same relation pattern, and the sample spatial distance between the sample node and the second-order neighbor node in the relation path, are input into the spatial relation perception model. The spatial relation perception layer in the spatial relation perception model outputs the third sample POI representation information of the sample node under the relation path. The first fusion layer in the spatial relationship perception model performs a fusion operation on all the third sample POI representation information of the sample node under the same relationship pattern to obtain the second sample POI representation information of the sample node under the same relationship pattern. The second fusion layer in the target spatial relationship perception model performs a fusion operation on the second sample POI representation information of the sample node in each relationship mode to obtain the first sample POI representation information of the sample node in all relationship modes.

14. The method according to claim 13, wherein, The step of inputting each relation path in the relation subset corresponding to the same relation pattern, and the spatial distance between nodes in the relation path and their second-order neighbor nodes, into the target spatial relation perception model, yields the third POI representation information of the node under the relation path, including: The first linear transformation matrix of the spatial relationship perception layer is used to linearly transform the first sample feature vector of the first-order neighbor node and the second sample feature vector of the second-order neighbor node of the sample node to determine the sample relationship weight vector of the sample node under the relationship path. The sample representation vector and the second sample feature vector of the sample spatial distance are linearly transformed by the first linear transformation matrix and the second linear transformation matrix of the spatial relationship perception layer to determine the sample spatial weight vector of the sample node under the relationship path; The sample space weight vector and the sample relationship weight vector are fused, and the fused sample vector is multiplied with the first model parameter of the spatial relationship perception layer and then mapped by the activation function to obtain the sample weight coefficient of the second-order neighbor node. Based on the first linear transformation matrix and the sample weight coefficients, the first sample feature vector and the second sample feature vector are fused to obtain the third sample POI representation information of the sample node under the relation path.

15. The method according to claim 14, wherein, The process of fusing the first sample feature vector and the second sample feature vector based on the first linear transformation matrix and the sample weight coefficients to obtain the third sample POI representation information of the sample node under the relation path includes: Based on the first linear transformation matrix, the feature vector of the first sample is linearly transformed to obtain the first transformation result; The second sample feature vector and the sample weight coefficient are weighted to obtain a weighted second sample feature vector. The weighted second sample feature vector is then linearly transformed based on the first linear transformation matrix to obtain a second transformation result. The first transformation result and the second transformation result are added together to obtain the third sample POI representation information of the sample node under the relation path.

16. The method according to claim 14 or 15, wherein, The process of refining the spatial relationship perception model based on the downstream task prediction data and the downstream task annotation data of the sample nodes until the training is completed and the target spatial relationship perception model is obtained includes: Based on the downstream task prediction data and the downstream task annotation data of the sample nodes, the loss function of the spatial relationship perception model is determined. Based on the loss function, the first linear transformation matrix, the second linear transformation matrix, and the first model parameters in the spatial relationship perception model are corrected, and the next sample is returned to continue training the corrected spatial relationship perception model until the training ends and the target spatial relationship perception model is obtained. The target spatial relationship perception model includes the first target linear transformation matrix, the second target linear transformation matrix, and the first target model parameters.

17. A device for acquiring representation information of points of interest (POIs), comprising: The first acquisition module is used to determine the POI relationship graph based on the map, and based on the POI relationship graph, acquire the node relationship set of each node on the POI relationship graph, wherein the POI relationship graph includes nodes representing POIs and edges of relationships between nodes, and the node relationship set includes relationship paths from a node to all second-order neighbor nodes via any first-order neighbor node. The second acquisition module is used to acquire the spatial distance matrix between the node and its corresponding second-order neighbor nodes, wherein the spatial distance matrix is ​​determined based on the spatial position of the node; The information acquisition module is used to determine the first POI representation information of the node based on the node relationship set and the spatial distance matrix. The first POI representation information of the node is used to perform downstream task processing on the node to obtain the processing result of the downstream task.

18. The apparatus according to claim 17, wherein, The information acquisition module includes: From the set of node relationships, determine a subset of relationships for the node under different relationship patterns, wherein the subset of relationships includes at least one relationship path for the node under the same relationship pattern; For each of the relation patterns, based on the relation paths in the relation subset and the spatial distance matrix, the second POI representation information of the node under the relation pattern is determined; The second POI representation information of the node in each relation schema is fused to obtain the first POI representation information of the node in all relation schemas.

19. The apparatus according to claim 18, wherein, The information acquisition module includes: For each relation path in the relation subset corresponding to the same relation schema, the spatial distance between the node in the relation path and its second-order neighbor node is determined based on the spatial distance matrix. Based on the spatial distance and the first-order and second-order neighbor nodes of the node in the relation path, determine the third POI representation information of the node under the relation path; The third POI representation information of the node under the same relation schema is fused to obtain the second POI representation information of the node under the same relation schema.

20. The apparatus according to claim 19, wherein, The information acquisition module includes: Based on the first-order neighbor node and the second-order neighbor node, determine the relation weight vector of the node under the relation path; Based on the spatial distance and the second-order neighbor nodes, determine the spatial weight vector of the node under the relation path; The spatial weight vector and the relational weight vector are fused to obtain the weight coefficients of the second-order neighbor nodes; Based on the first-order neighbor node, the second-order neighbor node, and the weight coefficient, the third POI representation information of the node under the relation path is obtained.

21. The apparatus according to claim 20, wherein, The information acquisition module includes: Determine the first feature vector of the first-order neighbor node and the second feature vector of the second-order neighbor node of the node in the relation path; Perform a linear transformation on the first feature vector and the second feature vector to determine the relation weight vector corresponding to the node.

22. The apparatus according to claim 20, wherein, The information acquisition module includes: Determine the representation vector of the spatial distance and the second feature vector of the second-order neighbor nodes; A linear transformation is performed on the spatial distance representation vector and the second feature vector to determine the spatial weight vector corresponding to the node.

23. The apparatus according to claim 17, wherein, The information acquisition module includes: Obtain a pre-trained target spatial relationship perception model; Each relation path in the relation subset corresponding to the same relation pattern, and the spatial distance between the node in the relation path and the second-order neighbor node, are input into the target spatial relation perception model. The spatial relation perception layer in the target spatial relation perception model outputs the third POI representation information of the node under the relation path. The first fusion layer in the target spatial relationship perception model performs a fusion operation on all the third POI representation information of the node under the same relationship pattern to obtain the second POI representation information of the node under the same relationship pattern. The second fusion layer in the target spatial relationship perception model performs a fusion operation on the second POI representation information of the node in each relationship mode to obtain the first POI representation information of the node in all relationship modes.

24. The apparatus according to claim 23, wherein, The information acquisition module includes: The first target linear transformation matrix of the spatial relationship perception layer is used to linearly transform the first feature vector of the first-order neighbor node and the second feature vector of the second-order neighbor node of the node to determine the relationship weight vector of the node under the relationship path. The spatial distance representation vector and the second feature vector are linearly transformed using the first target linear transformation matrix and the second target linear transformation matrix of the spatial relationship perception layer to determine the spatial weight vector of the node under the relationship path; The spatial weight vector and the relation weight vector are fused, and the fused vector is multiplied with the first target model parameters of the spatial relation perception layer and then mapped by an activation function to obtain the weight coefficients of the second-order neighbor nodes. Based on the first target linear transformation matrix and the weight coefficients, the first feature vector and the second feature vector are fused to obtain the third POI representation information of the node under the relation path.

25. The apparatus according to claim 24, wherein, The information acquisition module includes: A first transformation result is obtained by performing a linear transformation on the first feature vector based on the first target linear transformation matrix. The second feature vector and the weight coefficient are weighted to obtain a weighted second feature vector, and a linear transformation is performed on the weighted second feature vector based on the first target linear transformation matrix to obtain a second transformation result; The first transformation result and the second transformation result are added together to obtain the third POI representation information of the node under the relationship path.

26. The apparatus according to any one of claims 17-25, wherein, The first acquisition module includes: For the node v i any candidate second-order neighbor node v k Based on the POI relationship graph, obtain the nodes v i Passing through any first-order neighbor node v j Reaching the node v k The number of paths in the relational path; From the node v i Among all candidate second-order neighbor nodes, one or more candidate second-order neighbor nodes whose number of paths is greater than a set number are identified and retained as the target second-order neighbor nodes of the node. The relationship paths between the node and the target second-order neighbor node are mined to obtain the node relationship set of the node.

27. The apparatus according to any one of claims 17-25, wherein, The second acquisition module includes: For any second-order neighbor node of the node, obtain the first spatial coordinate information of the node and the second spatial coordinate information of the second-order neighbor node; Euclidean distance calculation is performed based on the first spatial coordinate information and the second spatial coordinate information to obtain the spatial distance between the node and its second-order neighbor nodes. The spatial distance matrix is ​​obtained based on the spatial distance between the node and each of its second-order neighbor nodes.

28. A training device for a spatial relationship-aware model of points of interest (POIs), comprising: The first sample acquisition module is used to determine the sample POI relationship graph based on the map, and based on the sample POI relationship graph, to obtain the sample node relationship set of each sample node on the sample POI relationship graph. The sample POI relationship graph includes sample nodes representing POIs and edges of relationships between sample nodes. The sample node relationship set includes the relationship paths of sample nodes reaching all second-order neighbor nodes through any first-order neighbor node. The second sample acquisition module is used to acquire the sample spatial distance matrix between the sample node and its corresponding second-order neighbor node, wherein the spatial distance matrix is ​​determined based on the spatial position of the sample node. The sample information acquisition module is used to train an initial spatial relationship perception model based on the sample node relationship set and the sample spatial distance matrix of the sample node, and to determine the first sample POI representation information of the sample node. The first POI representation information of the sample node is used to process the sample node for downstream tasks and obtain the processing result of the downstream tasks. The correction module is used to obtain the downstream task prediction data of the sample node based on the first sample POI representation information, and to correct the spatial relationship perception model based on the downstream task prediction data and the downstream task annotation data of the sample node until the training ends and the target spatial relationship perception model is obtained.

29. The apparatus according to claim 28, wherein, The sample information acquisition module includes: Each relation path in the relation subset corresponding to the same relation pattern, and the sample spatial distance between the sample node and the second-order neighbor node in the relation path, are input into the spatial relation perception model. The spatial relation perception layer in the spatial relation perception model outputs the third sample POI representation information of the sample node under the relation path. The first fusion layer in the spatial relationship perception model performs a fusion operation on all the third sample POI representation information of the sample node under the same relationship pattern to obtain the second sample POI representation information of the sample node under the same relationship pattern. The second fusion layer in the target spatial relationship perception model performs a fusion operation on the second sample POI representation information of the sample node in each relationship mode to obtain the first sample POI representation information of the sample node in all relationship modes.

30. The apparatus according to claim 29, wherein, The sample information acquisition module includes: The first linear transformation matrix of the spatial relationship perception layer is used to linearly transform the first sample feature vector of the first-order neighbor node and the second sample feature vector of the second-order neighbor node of the sample node to determine the sample relationship weight vector of the sample node under the relationship path. The sample representation vector and the second sample feature vector of the sample spatial distance are linearly transformed by the first linear transformation matrix and the second linear transformation matrix of the spatial relationship perception layer to determine the sample spatial weight vector of the sample node under the relationship path; The sample space weight vector and the sample relationship weight vector are fused, and the fused sample vector is multiplied with the first model parameter of the spatial relationship perception layer and then mapped by the activation function to obtain the sample weight coefficient of the second-order neighbor node. Based on the first linear transformation matrix and the sample weight coefficients, the first sample feature vector and the second sample feature vector are fused to obtain the third sample POI representation information of the sample node under the relation path.

31. The apparatus according to claim 30, wherein, The sample information acquisition module includes: Based on the first linear transformation matrix, the feature vector of the first sample is linearly transformed to obtain the first transformation result; The second sample feature vector and the sample weight coefficient are weighted to obtain a weighted second sample feature vector. The weighted second sample feature vector is then linearly transformed based on the first linear transformation matrix to obtain a second transformation result. The first transformation result and the second transformation result are added together to obtain the third sample POI representation information of the sample node under the relation path.

32. The apparatus according to claim 30 or 31, wherein, The correction module includes: Based on the downstream task prediction data and the downstream task annotation data of the sample nodes, the loss function of the spatial relationship perception model is determined. Based on the loss function, the first linear transformation matrix, the second linear transformation matrix, and the first model parameters in the spatial relationship perception model are corrected, and the next sample is returned to continue training the corrected spatial relationship perception model until the training ends and the target spatial relationship perception model is obtained. The target spatial relationship perception model includes the first target linear transformation matrix, the second target linear transformation matrix, and the first target model parameters.

33. An electronic device, comprising: At least one processor; as well as A memory communicatively connected to the at least one processor; wherein, The memory stores instructions executable by the at least one processor, which, when executed by the at least one processor, enables the at least one processor to perform the method of any one of claims 1-11 or 12-16.

34. A non-transitory computer-readable storage medium storing computer instructions, wherein, The computer instructions are used to cause the computer to perform the method according to any one of claims 1-11 or 12-16.

35. A computer program product comprising a computer program that, when executed by a processor, implements the steps of the method according to any one of claims 1-11 or 12-16.

Images