Point of interest recommendation method, system and equipment based on spatio-temporal context awareness

Abstract

The invention relates to the technical field of electric data processing, in particular to an interest point recommendation method, system and equipment based on spatio-temporal context awareness, and aims to solve the technical problem of low accuracy of a recommendation method in the prior art. The obtaining module is used for obtaining target user features containing social information; then, based on the target user, the friend user and the plurality of interest points, constructing a user interest point interaction bipartite graph for obtaining target user features containing interest point interaction preferences and a plurality of interest point interaction preference features; then, constructing an interest point space-time transfer diagram based on time and space information of accessing the plurality of interest points by the target user to obtain space-time preference characteristics of the plurality of interest points; and finally, performing probability mapping processing on the features to obtain recommendation values, and recommending the corresponding interest points before the sorting value of the recommendation values is smaller than a sorting threshold value to the target user, so that the recommendation accuracy is improved, and the user experience is improved.

Description

Technical Field

[0001] The present invention relates to the field of electronic data processing technology, and in particular to a method, system and device for recommending points of interest based on spatiotemporal context perception. Background Art

[0002] With the rapid development of Internet technology, social networking services continue to emerge in daily life, the number of users has increased rapidly, and a large amount of check-in data has been accumulated. In these networks, users can use location services with the help of smartphones and other mobile devices to explore nearby points of interest, record personal visit experiences, and share real-life experiences with others. These platforms not only promote interaction between users, but also greatly enrich people's lifestyles. As an extension of general point of interest recommendation research, next point of interest recommendation has become the research focus in the current service recommendation field.

[0003] Early POI recommendations mainly relied on traditional technologies such as content-based recommendations and collaborative filtering. In recent years, with the rapid development of advanced technologies such as machine learning, deep learning, natural language processing, and knowledge graphs, these methods have shown significant advantages in many fields. However, these methods usually only focus on shallow interactions between features and cannot obtain the deeper implicit relationship between users and POIs, resulting in low accuracy of service recommendations, reduced user interest, and affected user experience. Summary of the invention

[0004] The purpose of the present invention is to provide a method, system and device for recommending points of interest based on spatiotemporal context perception.

[0005] The technical solution of the present invention is as follows:

[0006] A method for recommending points of interest based on spatiotemporal context awareness includes the following operations:

[0007] S1. Obtain users who have social connections with the target user as friend users; construct a user social network graph with all users as nodes and edges as social connection strengths; transfer the friend user features in the user social network graph to the target user based on the multi-head attention mechanism to obtain the target user features containing social information;

[0008] S2. Based on the target user, friend users and several points of interest, a bipartite graph of user point of interest interaction is constructed; in the bipartite graph of user point of interest interaction, the edges between users and the edges between users and points of interest are obtained based on the interaction information between users and points of interest, and the edges between points of interest are obtained based on the popularity difference between points of interest; the bipartite graph of user point of interest interaction is subjected to several node information mutual propagation and node propagation information aggregation processes to obtain a user point of interest interaction propagation graph; the node features corresponding to the target user and several points of interest in the user point of interest interaction propagation graph are respectively subjected to nonlinear mapping processes to obtain target user features containing point of interest interaction preferences, as well as several point of interest interaction preference features;

[0009] S3. Based on the time and space information of the target user visiting several points of interest within the historical neighborhood time range, a spatiotemporal transfer graph of points of interest is constructed with points of interest as nodes and edges as the spatiotemporal intensity of visits; in the interest point transfer preference graph, all nodes are respectively subjected to several graph convolutions and feature update processes within the neighborhood range to obtain a spatiotemporal feature aggregation graph of points of interest; the corresponding node features of several points of interest in the spatiotemporal feature aggregation graph of points of interest are obtained to obtain several spatiotemporal preference features of points of interest;

[0010] S4. The target user features containing social information are fused with the target user features containing interest point interaction preferences to obtain target user fusion features; several interest point interaction preference features are fused with their corresponding interest point spatiotemporal preference features to obtain several interest point fusion features; the target user fusion features are probability mapped with several interest point fusion features to obtain several recommendation values; the several recommendation values ​​are arranged in descending order, and the interest points corresponding to the ranking values ​​less than the ranking threshold are recommended to the target user.

[0011] The operation of obtaining the target user features containing social information in S1 is implemented by the following formula:

[0012]

[0013] Q u is the target user feature containing social information, w k is the weight of the kth attention head, K is the total number of attention heads, α u,v is the social connection strength between target user u and friend user v, N is the user set, q v is the information feature of the vth friend user.

[0014] The strength of social connections between users in S1 is obtained by multiplying the information features of both users by their respective weights, and then performing concatenation, feature mapping, nonlinear processing, and normalization.

[0015] In the user interest point interaction bipartite graph of S2, if the user has visited an interest point, the edge between the user and the interest point is a positive edge; if the user has not visited the interest point, the edge between the user and the interest point is a negative edge; if the edges between users have the same sign as the same interest point, the edge between users is a positive edge; if there is no edge with the same sign as the same interest point between users, the edge between users is a negative edge; the edge between interest points is a vector edge, pointing from the first interest point to the second interest point, and the popularity of the first interest point is not less than that of the second interest point.

[0016] The popularity of a point of interest is obtained based on the number of times the point of interest has been visited by users within the historical neighborhood time range.

[0017] The operations of mutual propagation of current node information and aggregation processing of node propagation information in S2 are implemented by the following formulas:

[0018]

[0019] is the propagation characteristic from node x to node y for the l+1th time, β x,y is the weight corresponding to the positive edge, negative edge or vector edge between node x and node y, is the information aggregation feature of the l-th node x; is the aggregation coefficient of node x and node o at the lth time, which is obtained by multiplying the information aggregation features of node x at the l-1th time and the information aggregation features of node o at the l-1th time with their respective weights, and then performing concatenation, feature mapping, nonlinear processing and normalization; is the propagation characteristic from node o to node x for the lth time, and O is the set of all nodes.

[0020] In S3, the operation of graph convolution within the l+1th neighborhood is implemented by the following formula: is the graph convolution feature of the i-th node at the l+1th time, J is the set of nodes in the neighborhood of the i-th node, and A ij is the spatiotemporal adjacency matrix of the graph structure composed of nodes in the neighborhood of the i-th node, c is the normalization coefficient, W l +1 is the l+1th graph convolution weight, The feature update matrix of the jth node is updated for the lth time; the feature update processing within the lth neighborhood is implemented by the following formula:

[0021]

[0022] Update the matrix for the feature of the i-th node at the lth time, is the update coefficient of the i-th node at the lth time, Update the state value of the i-th node for the l-1th time, ⊙ is the Hadamard product operation, Update the state value of the i-th node for the l-th time, σ is the sigmoid function, W z Update the learning weight for the first node, U z Update the learning weight for the second node, b is the updated compensation value, is the graph convolution feature of the i-th node at the lth time, W a Update the learned weights for the nodes.

[0023] A point of interest recommendation system based on spatiotemporal context awareness, used to implement the above-mentioned point of interest recommendation method based on spatiotemporal context awareness, comprising:

[0024] The target user feature generation module containing social information is used to obtain users who have social connections with the target user as friend users; construct a user social network graph with all users as nodes and edges as social connection strengths; based on the multi-head attention mechanism, the friend user features in the user social network graph are transferred to the target user to obtain the target user features containing social information;

[0025] The module for generating target user features and interest point interaction preference features including interest point interaction preference is used to construct a user interest point interaction bipartite graph based on the target user, friend users and several interest points; in the user interest point interaction bipartite graph, the edges between users and the edges between users and interest points are obtained based on the interaction information between users and interest points, and the edges between interest points are obtained based on the popularity difference between interest points; the user interest point interaction bipartite graph is subjected to several node information mutual propagation and node propagation information aggregation processing to obtain a user interest point interaction propagation graph; the node features corresponding to the target user and several interest points in the user interest point interaction propagation graph are respectively subjected to nonlinear mapping processing to obtain the target user features including interest point interaction preference and several interest point interaction preference features;

[0026] The module for generating spatiotemporal preference features of interest points is used to construct a spatiotemporal transfer graph of interest points with interest points as nodes and edges as the spatiotemporal intensity of visits based on the temporal and spatial information of the target user visiting several interest points within the historical neighborhood time range; in the interest point transfer preference graph, all nodes are respectively subjected to several graph convolutions and feature update processes within the neighborhood range to obtain a spatiotemporal feature aggregation graph of interest points; the node features corresponding to several interest points in the spatiotemporal feature aggregation graph of interest points are obtained to obtain several spatiotemporal preference features of interest points;

[0027] The recommendation module is used to fuse the target user features containing social information with the target user features containing interest point interaction preferences to obtain the target user fusion features; several interest point interaction preference features are respectively fused with their corresponding interest point spatiotemporal preference features to obtain several interest point fusion features; the target user fusion features and several interest point fusion features are respectively subjected to probability mapping processing to obtain several recommendation values; the several recommendation values ​​are arranged in order from large to small, and the interest points corresponding to the sorting values ​​less than the sorting threshold are recommended to the target user.

[0028] A device for recommending points of interest based on spatiotemporal context awareness comprises a processor and a memory, wherein the processor implements the above-mentioned method for recommending points of interest based on spatiotemporal context awareness when executing a computer program stored in the memory.

[0029] A computer-readable storage medium is used to store a computer program, wherein when the computer program is executed by a processor, the above-mentioned method for recommending points of interest based on spatiotemporal context awareness is implemented.

[0030] The beneficial effects of the present invention are:

[0031] The present invention provides a method for recommending points of interest based on spatiotemporal context awareness. First, friend users who have social connections with a target user are obtained, and a user social network graph is constructed together with the target user. The friend user features in the user social network graph are transferred to the target user based on a multi-head attention mechanism, and the corresponding node features of the target user are extracted to obtain the target user features containing social information. Then, based on the target user, the friend user and a number of points of interest, a user point of interest interaction bipartite graph that can intuitively reflect the user's point of interest interaction preference and the point of interest popularity difference information is constructed. After several times of mutual propagation of node information and aggregation of node propagation information, nonlinear mapping processing is performed to obtain the target user features containing the point of interest interaction preference and a number of point of interest interaction preference features. Then, based on the time when the target user visits a number of points of interest, a target user feature containing the point of interest interaction preference and a number of point of interest interaction preference features are obtained. The temporal and spatial information of the target user is used to construct a spatiotemporal transfer graph of interest points that can intuitively reflect the spatiotemporal transfer preference of the target user's interest points. After several graph convolutions and feature updates within the neighborhood, the complex spatiotemporal transfer relationship between interest points is captured, and the contextual information of the target user's interest point transfer preference is extracted to obtain several spatiotemporal preference features of interest points. Finally, the target user features containing social information are probabilistically mapped with the target user features containing interest point interaction preferences, several interest point interaction preference features, and several interest point spatiotemporal preference features to obtain the recommended value, and the interest points corresponding to the recommended value before the ranking value is less than the ranking threshold are recommended to the target user, thereby improving the recommendation accuracy and user experience. Compared with the existing methods, the method of the present invention can improve the accuracy and recall rate, with an average improvement rate of 24.22% and 12.53%. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] By reading the detailed description of the preferred embodiment below, the scheme and advantages of the present application will become clear to those skilled in the art. The accompanying drawings are only for the purpose of illustrating the preferred embodiment and are not to be considered as limiting the present invention.

[0033] In the attached picture:

[0034] Figure 1 In the embodiment, a schematic diagram of the process of the method of this embodiment;

[0035] Figure 2 Schematic diagram of the structure of the spatiotemporal transfer graph of interest points in the embodiment. DETAILED DESCRIPTION

[0036] Exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings.

[0037] This embodiment provides a method for recommending points of interest based on spatiotemporal context awareness. Figure 1 , including the following operations:

[0038] S1. Obtain users who have social connections with the target user as friend users; construct a user social network graph with all users as nodes and edges as social connection strengths; transfer the friend user features in the user social network graph to the target user based on the multi-head attention mechanism to obtain the target user features containing social information;

[0039] S2. Based on the target user, friend users and several points of interest, a bipartite graph of user point of interest interaction is constructed; in the bipartite graph of user point of interest interaction, the edges between users and the edges between users and points of interest are obtained based on the interaction information between users and points of interest, and the edges between points of interest are obtained based on the popularity difference between points of interest; the bipartite graph of user point of interest interaction is subjected to several node information mutual propagation and node propagation information aggregation processes to obtain a user point of interest interaction propagation graph; the node features corresponding to the target user and several points of interest in the user point of interest interaction propagation graph are respectively subjected to nonlinear mapping processes to obtain target user features containing point of interest interaction preferences, as well as several point of interest interaction preference features;

[0040] S3. Based on the time and space information of the target user visiting several points of interest within the historical neighborhood time range, a spatiotemporal transfer graph of points of interest is constructed with points of interest as nodes and edges as the spatiotemporal intensity of visits; in the interest point transfer preference graph, all nodes are respectively subjected to several graph convolutions and feature update processes within the neighborhood range to obtain a spatiotemporal feature aggregation graph of points of interest; the corresponding node features of several points of interest in the spatiotemporal feature aggregation graph of points of interest are obtained to obtain several spatiotemporal preference features of points of interest;

[0041] S4. The target user features containing social information are fused with the target user features containing interest point interaction preferences to obtain target user fusion features; several interest point interaction preference features are fused with their corresponding interest point spatiotemporal preference features to obtain several interest point fusion features; the target user fusion features are probability mapped with several interest point fusion features to obtain several recommendation values; the several recommendation values ​​are arranged in descending order, and the interest points corresponding to the ranking values ​​less than the ranking threshold are recommended to the target user.

[0042] S1. Obtain users who have social connections with the target user as friend users; construct a user social network graph with all users as nodes and edges as social connection strengths; transfer the friend user features in the user social network graph to the target user based on the multi-head attention mechanism to obtain the target user features containing social information.

[0043] Acquire friend users who have social connections with the target user, build a user social network graph with the target user, and transfer the friend user features in the user social network graph to the target user based on the multi-head attention mechanism, extract the corresponding node features of the target user, and obtain the target user features containing social information, so as to obtain the target user's social tendencies, indirectly obtain the target user's preference information, increase the target user's preference information content, solve the problem of sparse target user data, and improve the accuracy of subsequent interest point recommendations.

[0044] First, users who have social connections with the target user are obtained as friend users. Based on the social connections between the target user and the friend users, the target user's social tendencies are obtained, thereby indirectly obtaining the target user's preference information and increasing the target user's preference information content.

[0045] Then, a user social network graph is constructed with all users (including target users and friend users) as nodes and edges as social connection strengths, so as to intuitively reflect the user's social tendencies.

[0046] Finally, based on the multi-head attention mechanism, the friend user features in the user social network graph are transferred to the target user, the corresponding node features of the target user are extracted, and the target user features containing social information are obtained.

[0047] The above operation of obtaining the target user characteristics containing social information is achieved through the following formula:

[0048]

[0049] Q u is the target user feature containing social information, w k is the weight of the kth attention head, K is the total number of attention heads, α u,vis the social connection strength between the target user u and the friend user v, N is the user set, and the user set N includes the target user u, α u,u =1,q v is the information feature of the vth friend user.

[0050] The strength of social connections between users is obtained by multiplying the information features of both users by their respective weights, and then performing splicing, feature mapping, nonlinear processing and normalization.

[0051] The social connection strength is obtained by the following formula:

[0052]

[0053] w u is the weight of target user u, q u is the information feature of the target user u, w v is the weight of target user v, q v is the information feature of friend user v, a() is the feature mapping function, which maps the concatenated high-dimensional features to a real number (which can be achieved through convolutional layers and pooling layers), and LeakyReLU() is a nonlinear function.

[0054] S2. Based on the target user, friend users and several interest points, a bipartite graph of user interest point interaction is constructed; the bipartite graph of user interest point interaction is subjected to several node information mutual propagation and node propagation information aggregation processes to obtain a user interest point interaction propagation graph; in the user interest point interaction propagation graph, the target user and several interest points corresponding node features are subjected to nonlinear mapping processes to obtain target user features containing interest point interaction preferences, as well as several interest point interaction preference features.

[0055] Based on the target user, friend users and several interest points, a bipartite graph of user interest point interaction is constructed, which can intuitively reflect the user's interest point interaction preference and interest point popularity difference information. After several rounds of node information mutual propagation and node propagation information aggregation processing, nonlinear mapping processing is performed to obtain the target user characteristics containing interest point interaction preference and several interest point interaction preference characteristics.

[0056] First, based on the target user, friend users and several points of interest to be recommended, a bipartite graph of user point of interest interaction is constructed to more comprehensively understand the target user's point of interest interaction preferences. In the bipartite graph of user point of interest interaction, the edges between users and the edges between users and points of interest are obtained based on the interaction information between users and points of interest, and the edges between points of interest are obtained based on the popularity difference between points of interest.

[0057] The edge attributes between users and points of interest are defined as follows: if the user has visited a point of interest, it means that the user preference is closely related to the point of interest, and the edge between the user and the point of interest is a positive edge; if the user has not visited a point of interest, it means that the user preference is not closely related to the point of interest, and the edge between the user and the point of interest is a negative edge.

[0058] The edge attributes between users are defined as follows: if the edge signs between users are the same as the same point of interest, the edge between users is a positive edge; if the edge signs between users do not have the same point of interest, the edge between users is a negative edge. For example, if the edges between user u and user v and point of interest i are both positive or negative, it means that user u and user v are both interested in or not interested in point of interest i, which means that user u and user v have similar preferences, so the edge between user u and user v is a positive edge, otherwise it is a negative edge.

[0059] The edge attributes between interest points are defined as follows: the edge between interest points is a vector edge, pointing from the first interest point to the second interest point, and the popularity of the first interest point is not less than that of the second interest point. For example, the popularity of interest point i is 6, and the popularity of interest point j is 4. The edge between interest point i and interest point j is a vector edge with an arrow pointing from interest point i to interest point j, reflecting the decreasing direction of popularity, which is convenient for analyzing the popularity distribution of interest points. The vector edge weight is obtained based on the popularity difference between interest points, and the popularity of interest points is obtained based on the number of times the interest points are visited by users within the historical neighborhood time range.

[0060] Then, the user interest point interaction bipartite graph is subjected to several node information mutual propagation and node propagation information aggregation processes to comprehensively learn the interaction preferences of interest points and obtain the user interest point interaction propagation graph.

[0061] The current (l+1th) node information propagation is achieved through the following formula:

[0062]

[0063] is the propagation characteristic from node x to node y for the l+1th time, β x,y is the weight corresponding to the positive edge, negative edge or vector edge between node x and node y. The weight corresponding to the edge between users is obtained based on the edge attribute and / or the strength of social connection between users and / or the total number of common interactive interest points. The weight corresponding to the edge between users and interest points is obtained based on the edge attribute and / or the number of interactions. The vector edge weight is obtained based on the popularity difference between interest points. is the information aggregation feature of the l-th node x.

[0064] The operations of the l-th node propagation information aggregation processing are implemented by the following formulas:

[0065]

[0066] is the aggregation coefficient of node x and node o at the lth time, which is obtained by multiplying the information aggregation features of node x at the l-1th time and the information aggregation features of node o at the l-1th time with their respective weights, and then performing concatenation, feature mapping, nonlinear processing and normalization; is the propagation feature of node o to node x for the lth time, including the propagation feature from user node to interest point node, the propagation feature from user node to user node, and the propagation feature from the first interest point node to the second interest point node. O is the set of all nodes, and the node set O includes node x.

[0067] Finally, the target user and several corresponding node features of interest points in the user interest point interaction propagation graph are processed nonlinearly to obtain the target user features containing interest point interaction preferences and several interest point interaction preference features.

[0068] The nonlinear mapping process is implemented by the following formula:

[0069] R u =W r (σ(dropout(W1p u +d1)))+d2,

[0070] R u is the target user feature with interest point interaction preference, W r is the nonlinear mapping processing weight, σ is the sigmoid function, dropout is the parameter loss function, p u is the node feature (information aggregation feature) corresponding to the target user u, d1 is the first nonlinear mapping compensation, and d2 is the second nonlinear mapping compensation.

[0071] S3. Based on the temporal and spatial information of the target user visiting several points of interest within the historical neighborhood time range, a spatiotemporal transfer graph of points of interest is constructed with points of interest as nodes and edges as the spatiotemporal intensity of visits; in the interest point transfer preference graph, all nodes are subjected to graph convolution and feature update processing within the neighborhood several times to obtain a spatiotemporal feature aggregation graph of points of interest; the corresponding node features of several points of interest in the spatiotemporal feature aggregation graph of points of interest are obtained to obtain several spatiotemporal preference features of points of interest.

[0072] Based on the temporal and spatial information of the target user's visits to several points of interest, a spatiotemporal transfer graph of interest points is constructed, which can intuitively reflect the spatiotemporal transfer preference of the target user's interest points. After several graph convolutions and feature updates within the neighborhood, the complex spatiotemporal transfer relationship between interest points is captured, and the contextual information of the target user's interest point transfer preference is extracted to obtain several spatiotemporal preference features of interest points.

[0073] First, based on the temporal and spatial information of target users visiting several points of interest within the historical neighborhood time range, a spatiotemporal transfer graph of points of interest is constructed with points of interest as nodes and edges as the spatiotemporal intensity of visits, such as Figure 2 , which is convenient for capturing the spatial and temporal relationship of target user preferences and obtaining target user preference transfer information. The spatial-temporal intensity of visits is based on the spatial distance between nodes (points of interest) and the visit interval time.

[0074] Then, in the interest point transfer preference graph, all nodes are subjected to several graph convolution and feature update processes within the neighborhood range, and the node features containing similar target user preferences in time and space are aggregated to mine the key features reflecting the target user's interest point migration tendency, and obtain the interest point spatiotemporal feature aggregation graph.

[0075] The operation of graph convolution within the l+1th neighborhood is implemented by the following formula:

[0076]

[0077] is the graph convolution feature of the i-th node at the l+1th time, J is the set of nodes in the neighborhood of the i-th node, and A ij is the spatiotemporal adjacency matrix of the graph structure (the spatiotemporal transition graph of interest points) composed of the nodes in the neighborhood of the i-th node, which is obtained by the Hadamard product operation of the temporal adjacency matrix and the spatial adjacency matrix of the corresponding graph structure. The neighborhood range is obtained based on the preset spatiotemporal intensity of the visit; c is the normalization coefficient, W l+1 is the l+1th graph convolution weight, Update the matrix for the lth jth node feature.

[0078] The feature update operation within the lth neighborhood is implemented by the following formula:

[0079]

[0080] Update the matrix for the feature of the i-th node at the lth time, is the update coefficient of the i-th node at the lth time, Update the state value of the i-th node for the l-1th time, ⊙ is the Hadamard product operation, Update the state value of the i-th node for the l-th time, σ is the sigmoid function, W z Update the learning weight for the first node, U z Update the learning weight for the second node, b is the updated compensation value, is the graph convolution feature of the i-th node at the lth time, W a Update the learned weights for the nodes.

[0081] Finally, the corresponding node features of several interest points in the interest point spatiotemporal feature aggregation graph are obtained, and the relevant information about the target user's preference transfer on the interest point is extracted to obtain several interest point spatiotemporal preference features.

[0082] S4. The target user features containing social information are fused with the target user features containing interest point interaction preferences to obtain target user fusion features; several interest point interaction preference features are fused with their corresponding interest point spatiotemporal preference features to obtain several interest point fusion features; the target user fusion features are probability mapped with several interest point fusion features to obtain several recommendation values; the several recommendation values ​​are arranged in descending order, and the interest points corresponding to the ranking values ​​less than the ranking threshold are recommended to the target user.

[0083] The target user features containing social information are probabilistically mapped with the target user features containing interest point interaction preferences, several interest point interaction preference features, and several interest point spatiotemporal preference features to obtain recommendation values; and the interest points corresponding to the recommendation value before the ranking value is less than the ranking threshold are recommended to the target user to improve the recommendation accuracy.

[0084] Firstly, the target user features containing social information are fused with the target user features containing interest point interaction preferences to obtain the target user fusion features that can fully reflect the target user preferences.

[0085] At the same time, several interest point interaction preference features are fused with their corresponding interest point spatiotemporal preference features respectively to obtain several interest point fusion features that can reflect the interest point preference transfer of the target user.

[0086] Then, the target user fusion features and several interest point fusion features are respectively subjected to probability mapping processing (which can be achieved through the sigmoid function) to obtain several recommendation values.

[0087] Finally, several recommendation values ​​are arranged in order from large to small, and the points of interest corresponding to the sorting values ​​less than the sorting threshold, that is, the points of interest with larger recommendation values, are recommended to the target users to improve the recommendation accuracy and enhance the user experience.

[0088] To verify the effectiveness of the method recommended in this embodiment, the following experiment was conducted.

[0089] Dataset and experimental environment. Three sets of personalized POI recommendation data sets were selected in the experiment to verify the performance of the method of this embodiment, namely the Foursquare data set, the Gowalla data set, and the Yelp data set. These three data sets contain rich user social data and POI interaction data. Each piece of user POI interaction data records the user ID and POI information in detail. In order to ensure the accuracy and effectiveness of the experiment, and to be consistent with the experimental settings of the baseline model (existing method), the experiment referred to the existing method and performed data preprocessing operations on the original data set. In the data preprocessing stage, those POIs that were visited less than 10 times and users who checked in less than 10 times were removed. The details of the processed data set are shown in Table 1. Subsequently, 60% of each data set was randomly selected as training data, 20% as verification data, and the remaining 20% ​​was assigned as test data.

[0090] Table 1 Summary of data information of Foursquare dataset, Gowalla dataset and Yelp dataset

[0091] Dataset Number of users Number of points of interest Number of sign-ins Foursquare Dataset 2551 13474 124993 Gowalla Dataset 5628 31803 620683 Yelp Dataset 30887 18995 860888

[0092] The experimental operating environment is as follows: the GPU uses the NVIDIA GeForce RTX 3090 graphics card to ensure the efficiency of model training and reasoning; the operating system uses Windows 10 Professional Edition 64-bit to provide a stable computing environment; the memory configuration is 32GB RAM to meet the memory requirements of large neural network models; the CPU is the 12th generation Intel Core i9-12900K, with a main frequency of 3.19GHz, providing powerful computing power; the deep learning model framework uses Pytorch based on Python 3.8 version, which supports model building, training and deployment; the development environment uses the Pycharm integrated development environment to provide a convenient programming and debugging experience.

[0093] Experimental evaluation indicators. Three evaluation indicators were used in the experiment, namely, Precision (P@K), Recall (R@K) and MAP (M@K). The formulas of these three indicators are defined as follows:

[0094]

[0095] Among them, R(g) represents the list of POI recommendation results, T(g) is the real user selection interaction list, N is the length of the POI recommendation list, and the value of K is set at K = [5, 10, 20].

[0096] Comparison with existing methods. Seven state-of-the-art personalized POI recommendation models are as follows.

[0097] HME method: The HME method constructs a hyperspace embedding model that includes POI, user, region and category information, and uses the POI unit ball model to represent it. By jointly learning POI-POI transfer, user preferences, region relations and category relations, the HME method can effectively capture the implicit hierarchical structure. In order to combine user preferences with the sequential transfer between POI, the HME method designs the Einstein midpoint aggregation method to improve the recommendation accuracy.

[0098] CLB method: The CLB method proposes a new next point of interest recommendation system based on the user's current location. The model constructs a user's check-in sequence graph, represents the user's check-in sequence as a directed graph, and calculates the current location of similar users based on time and distance factors to generate a recommendation list.

[0099] SUCP method: The SUCP method combines the user's social network and the spatiotemporal data of the user's geographic activity center, and uses matrix decomposition methods to predict the user's preference for points of interest. At the same time, the model also considers multiple factors such as the similarity between the user and his friends, the user's check-in habits, and the geographical distribution of points of interest.

[0100] PPR_IGCN method: The PPR_IGCN method combines the user's social relationships and the spatiotemporal information of the user's geographic activity center to predict the user's interest point preferences, and comprehensively learns the user's interest preferences and behavior patterns through joint learning of multi-type relationship graphs.

[0101] RSTSE method: The RSTSE method constructs a virtual trajectory map that integrates the spatiotemporal characteristics of users and points of interest. It evaluates the similarity between users by calculating the direct and indirect trust between users. The model combines neural collaborative filtering technology to generate a recommended list of places that users may be interested in in the future.

[0102] RST-GCN method: The RST-GCN method effectively captures the dynamic interests of users by combining a multi-head self-attention mechanism and regionalized subgraphs. The model uses the user's check-in sequence, point of interest visit sequence, and geographic spatial features to classify users into different regional categories, reducing the information propagation between non-similar nodes; through the multi-head self-attention mechanism, the RST-GCN method dynamically adjusts the importance of each time step to better reflect the changes in user interests.

[0103] DeePOF method: The DeepPOF method comprehensively reviews the POI recommendation system based on deep learning, explores the application of various neural network methods in capturing user check-in sequences, POI visit sequences and geographic spatial features, and discusses the advantages of these methods in solving data sparsity and cold start problems.

[0104] The experiment trained and optimized the method of this embodiment on three public datasets: Gowalla dataset, Foursquare dataset and Yelp dataset. For the users in each dataset, the study predicted their ratings for all points of interest and sorted them according to these ratings. According to different values ​​of the number of recommendations K (K=5, K=10 and K=20), the study generated three different sets of prediction results. Then, these prediction results were compared and analyzed with the test dataset to evaluate the accuracy and recall of the model. In addition, these evaluation indicators were also used to compare the performance with the selected benchmark model. The experimental results are shown in Tables 2, 3 and 4. Since the evaluation indicator K values ​​of the various existing methods are not exactly the same as those of the method of this embodiment, data is missing in some cases. The missing data is marked with the " / " symbol in the table.

[0105] Experimental results on the Foursquare dataset. In the experiment, the HME method, CLB method, RST-GCN method, PPR_IGCN method and RSTSE method were selected as comparison methods with the method of this embodiment on the Foursquare dataset. From the experimental results in Table 2, it can be concluded that the method of this embodiment has an average improvement of 13.83% and 8.6% in the P@K and R@K indicators, respectively, and the overall performance of the method of this embodiment is better than other existing methods.

[0106] Table 2 Experimental comparison results of the method in this embodiment and the existing method on the Foursquare dataset

[0107]

[0108]

[0109] Experimental results on the Gowalla dataset. In the experiment, the HME method, RST-GCN method, CLB method, RSTSE method and SUCP method were selected as comparison methods with the method of this embodiment on the Gowalla dataset. From the experimental results in Table 3, it can be concluded that the method of this embodiment has an average improvement of 47% and 18% in the P@K and R@K indicators respectively, and the overall performance of the method of this embodiment is better than other existing methods.

[0110] Table 3 Experimental comparison results of the method in this embodiment and the existing methods on the Gowalla dataset

[0111]

[0112] Experimental results on the Yelp dataset. In the experiment, the SUCP method, PPR_IGCN method and DeepPOF method were selected as comparison models with the method of this embodiment on the Yelp dataset. From the experimental results in Table 4, it can be seen that the method of this embodiment has an average improvement of 11.83% and 11% in the P@K and R@K indicators respectively, and the overall performance of the method of this embodiment is better than other existing methods.

[0113] Table 4 Experimental comparison results of the method in this embodiment and the existing methods on the Yelp dataset

[0114]

[0115] In general, compared with existing advanced recommendation methods, the method of this embodiment shows significant advantages in both accuracy and recall, with average improvements of 24.22% and 12.53% respectively, and the overall performance of the method of this embodiment surpasses other existing methods.

[0116] This embodiment further provides a point of interest recommendation system based on spatiotemporal context awareness, which is used to implement the above-mentioned point of interest recommendation method based on spatiotemporal context awareness, including:

[0117] The target user feature generation module containing social information is used to obtain users who have social connections with the target user as friend users; construct a user social network graph with all users as nodes and edges as social connection strengths; based on the multi-head attention mechanism, the friend user features in the user social network graph are transferred to the target user to obtain the target user features containing social information;

[0118] The module for generating target user features and interest point interaction preference features including interest point interaction preference is used to construct a user interest point interaction bipartite graph based on the target user, friend users and several interest points; in the user interest point interaction bipartite graph, the edges between users and the edges between users and interest points are obtained based on the interaction information between users and interest points, and the edges between interest points are obtained based on the popularity difference between interest points; the user interest point interaction bipartite graph is subjected to several node information mutual propagation and node propagation information aggregation processing to obtain a user interest point interaction propagation graph; the node features corresponding to the target user and several interest points in the user interest point interaction propagation graph are respectively subjected to nonlinear mapping processing to obtain the target user features including interest point interaction preference and several interest point interaction preference features;

[0119] The module for generating spatiotemporal preference features of interest points is used to construct a spatiotemporal transfer graph of interest points with interest points as nodes and edges as the spatiotemporal intensity of visits based on the temporal and spatial information of the target user visiting several interest points within the historical neighborhood time range; in the interest point transfer preference graph, all nodes are respectively subjected to several graph convolutions and feature update processes within the neighborhood range to obtain a spatiotemporal feature aggregation graph of interest points; the node features corresponding to several interest points in the spatiotemporal feature aggregation graph of interest points are obtained to obtain several spatiotemporal preference features of interest points;

[0120] The recommendation module is used to fuse the target user features containing social information with the target user features containing interest point interaction preferences to obtain the target user fusion features; several interest point interaction preference features are respectively fused with their corresponding interest point spatiotemporal preference features to obtain several interest point fusion features; the target user fusion features and several interest point fusion features are respectively subjected to probability mapping processing to obtain several recommendation values; the several recommendation values ​​are arranged in order from large to small, and the interest points corresponding to the sorting values ​​less than the sorting threshold are recommended to the target user.

[0121] This embodiment also provides a device for recommending points of interest based on spatiotemporal context awareness, including a processor and a memory, wherein the processor implements the above-mentioned method for recommending points of interest based on spatiotemporal context awareness when executing a computer program stored in the memory.

[0122] This embodiment further provides a computer-readable storage medium for storing a computer program, wherein the computer program, when executed by a processor, implements the above-mentioned method for recommending points of interest based on spatiotemporal context awareness.

[0123] The present embodiment provides a method for recommending points of interest based on spatiotemporal context awareness. First, friend users who have social connections with a target user are obtained, and a user social network graph is constructed together with the target user. The friend user features in the user social network graph are transferred to the target user based on a multi-head attention mechanism, and the node features corresponding to the target user are extracted to obtain target user features containing social information. Then, based on the target user, the friend user and a number of points of interest, a user point of interest interaction bipartite graph that can intuitively reflect the user's point of interest interaction preference and the point of interest popularity difference information is constructed. After several times of mutual propagation of node information and aggregation of node propagation information, nonlinear mapping processing is performed to obtain target user features containing point of interest interaction preferences and a number of point of interest interaction preference features. Then, based on the target user's visits to a number of points of interest, a target user feature containing a point of interest interaction preference and a number of point of interest interaction preference features are obtained. The temporal and spatial information is used to construct a spatiotemporal transfer graph of interest points that can intuitively reflect the spatiotemporal transfer preference of the target user's interest points, and after several graph convolutions and feature update processes within the neighborhood, the complex spatiotemporal transfer relationship between interest points is captured, and the contextual information of the target user's interest point transfer preference is extracted to obtain several spatiotemporal preference features of interest points; finally, the target user characteristics containing social information are probabilistically mapped with the target user characteristics containing interest point interaction preferences, several interest point interaction preference features, and several interest point spatiotemporal preference features to obtain the recommended value, and the interest points corresponding to the recommended value ranking value less than the ranking threshold are recommended to the target user, thereby improving the recommendation accuracy and user experience; compared with the existing methods, the method of the present invention can improve the accuracy and recall rate, with an average improvement rate of 24.22% and 12.53%.

Claims

1. A method for recommending points of interest based on spatiotemporal context awareness, characterized in that: The following operations are included: S1. Obtain users who have social connections with the target user as friend users; construct a user social network graph with all users as nodes and edges as social connection strengths; transfer the friend user features in the user social network graph to the target user based on the multi-head attention mechanism to obtain the target user features containing social information; S2. Based on the target user, friend users and several interest points, a bipartite graph of user interest point interaction is constructed; in the bipartite graph of user interest point interaction, the edges between users and the edges between users and interest points are obtained based on the interaction information between users and interest points, and the edges between interest points are obtained based on the popularity difference between interest points; The user interest point interaction bipartite graph is processed by performing several node information mutual propagation and node propagation information aggregation processes to obtain the user interest point interaction propagation graph; the target user and several interest point corresponding node features in the user interest point interaction propagation graph are processed by nonlinear mapping respectively to obtain the target user features containing interest point interaction preference and several interest point interaction preference features; S3. Based on the time and space information of the target user visiting several points of interest within the historical neighborhood time range, a spatiotemporal transfer graph of points of interest is constructed with points of interest as nodes and edges as the spatiotemporal intensity of visits; in the interest point transfer preference graph, all nodes are respectively subjected to several graph convolutions and feature update processes within the neighborhood range to obtain a spatiotemporal feature aggregation graph of points of interest; the corresponding node features of several points of interest in the spatiotemporal feature aggregation graph of points of interest are obtained to obtain several spatiotemporal preference features of points of interest; S4, merging the target user features containing social information with the target user features containing interest point interaction preferences to obtain target user fusion features; Several interest point interaction preference features are fused with their corresponding interest point spatiotemporal preference features to obtain several interest point fusion features; The target user fusion features and several interest point fusion features are respectively subjected to probability mapping processing to obtain several recommendation values; the several recommendation values ​​are arranged in order from large to small, and the interest points corresponding to the ranking values ​​less than the ranking threshold are recommended to the target user.

2. The method for recommending points of interest based on spatiotemporal context awareness according to claim 1, characterized in that: In S1, the operation of obtaining the target user features containing social information is implemented by the following formula: Q u is the target user feature containing social information, w k is the weight of the kth attention head, K is the total number of attention heads, α u,v is the social connection strength between target user u and friend user v, N is the user set, q v is the information feature of the vth friend user.

3. The method for recommending points of interest based on spatiotemporal context awareness according to claim 2, characterized in that: In S1, the social connection strength between users is obtained by multiplying the information features of both users by their respective weights, and then performing concatenation, feature mapping, nonlinear processing and normalization.

4. The method for recommending points of interest based on spatiotemporal context awareness according to claim 1, characterized in that: In the user interest point interaction bipartite graph of S2, If the user has visited a point of interest, the edge between the user and the point of interest is a positive edge; if the user has not visited a point of interest, the edge between the user and the point of interest is a negative edge; If the edge between users has the same sign as the same point of interest, the edge between users is a positive edge; if the edge between users does not have the same sign as the same point of interest, the edge between users is a negative edge; The edge between the interest points is a vector edge, which points from the first interest point to the second interest point, and the popularity of the first interest point is not less than that of the second interest point.

5. The method for recommending points of interest based on spatiotemporal context awareness according to claim 4, characterized in that: The popularity of a point of interest is obtained based on the number of times the point of interest has been visited by users within the historical neighborhood time range.

6. The method for recommending points of interest based on spatiotemporal context awareness according to claim 1, characterized in that: In S2, the operations of mutual propagation of the current secondary node information and aggregation processing of node propagation information are respectively implemented by the following formulas: is the propagation characteristic from node x to node y for the l+1th time, β x,y is the weight corresponding to the positive edge, negative edge or vector edge between node x and node y, is the information aggregation feature of the l-th node x; is the aggregation coefficient of node x and node o for the lth time, which is obtained by multiplying the information aggregation features of node x for the l-1th time and the information aggregation features of node o for the l-1th time with their respective weights, and then performing concatenation, feature mapping, nonlinear processing and normalization; is the propagation characteristic from node o to node x for the lth time, and O is the set of all nodes.

7. The method for recommending points of interest based on spatiotemporal context awareness according to claim 1, characterized in that: In S3, The operation of graph convolution within the l+1th neighborhood is implemented by the following formula: is the graph convolution feature of the i-th node at the l+1th time, J is the set of nodes in the neighborhood of the i-th node, and A ij is the spatiotemporal adjacency matrix of the graph structure composed of nodes in the neighborhood of the i-th node, c is the normalization coefficient, W l+1 is the l+1th graph convolution weight, Update the matrix for the lth jth node feature; The feature update operation within the lth neighborhood is implemented by the following formula: Update the matrix for the feature of the i-th node at the lth time, is the update coefficient of the i-th node at the lth time, Update the state value of the i-th node for the l-1th time, ⊙ is the Hadamard product operation, Update the state value of the i-th node for the l-th time, σ is the sigmoid function, W z Update the learning weight for the first node, U z Update the learning weight for the second node, b is the updated compensation value, is the graph convolution feature of the i-th node at the lth time, W a Update the learned weights for the nodes.

8. A point of interest recommendation system based on spatiotemporal context awareness, used to implement the point of interest recommendation method based on spatiotemporal context awareness as claimed in claim 1, characterized in that: include: A target user feature generation module containing social information is used to obtain users who have social connections with the target user as friend users; Construct a user social network graph with all users as nodes and edges as social connection strength; transfer the friend user features in the user social network graph to the target user based on the multi-head attention mechanism to obtain the target user features containing social information; The module for generating target user features and interest point interaction preference features including interest point interaction preference is used to construct a user interest point interaction bipartite graph based on the target user, friend users and several interest points; in the user interest point interaction bipartite graph, the edges between users and the edges between users and interest points are obtained based on the interaction information between users and interest points, and the edges between interest points are obtained based on the popularity difference between interest points; the user interest point interaction bipartite graph is subjected to several node information mutual propagation and node propagation information aggregation processing to obtain a user interest point interaction propagation graph; the node features corresponding to the target user and several interest points in the user interest point interaction propagation graph are respectively subjected to nonlinear mapping processing to obtain the target user features including interest point interaction preference and several interest point interaction preference features; The module for generating spatiotemporal preference features of interest points is used to construct a spatiotemporal transfer graph of interest points with interest points as nodes and edges as the spatiotemporal intensity of visits based on the temporal and spatial information of the target user visiting several interest points within the historical neighborhood time range; in the interest point transfer preference graph, all nodes are respectively subjected to several graph convolutions and feature update processes within the neighborhood range to obtain a spatiotemporal feature aggregation graph of interest points; the node features corresponding to several interest points in the spatiotemporal feature aggregation graph of interest points are obtained to obtain several spatiotemporal preference features of interest points; The recommendation module is used to fuse the target user features containing social information with the target user features containing interest point interaction preferences to obtain the target user fusion features; Several interest point interaction preference features are fused with their corresponding interest point spatiotemporal preference features to obtain several interest point fusion features; the target user fusion feature and several interest point fusion features are probability mapped to obtain several recommendation values; the several recommendation values ​​are arranged in descending order, and the interest points corresponding to the ranking values ​​less than the ranking threshold are recommended to the target user.

9. A point of interest recommendation device based on spatiotemporal context perception, characterized in that: The method comprises a processor and a memory, wherein when the processor executes the computer program stored in the memory, the method for recommending points of interest based on spatiotemporal context awareness as described in any one of claims 1 to 7 is implemented.

10. A computer-readable storage medium, characterized in that: Used to store a computer program, wherein when the computer program is executed by a processor, the method for recommending points of interest based on spatiotemporal context awareness as described in any one of claims 1 to 7 is implemented.

Images