Point of interest recommendation method and apparatus, system for spatiotemporal modeling

Abstract

The present disclosure relates to a point-of-interest recommendation method and device, a spatiotemporal modeling system, and relates to the field of spatiotemporal modeling in artificial intelligence. The point-of-interest recommendation method comprises: dividing historical trajectories of a target user and other users into sessions respectively according to a specified time interval; generating a global interaction hypergraph according to the sessions, wherein the global interaction hypergraph comprises nodes and hyperedges, the nodes represent points of interest, and the hyperedges represent connections between points of interest in the same session; generating global point-of-interest information according to the global interaction hypergraph; and generating a target point of interest recommended to the target user according to initial information of the target user and the global point-of-interest information. According to the present disclosure, the accuracy of point-of-interest recommendation is improved.

Description

Technical Field

[0001] This disclosure relates to the field of spatiotemporal modeling in the field of artificial intelligence, and particularly to methods and apparatus for recommending points of interest, systems for spatiotemporal modeling, and computer-readable storage media. Background Technology

[0002] Points of Interest (POIs) are point-based data in internet electronic maps, typically containing attributes such as name, address, coordinates, and category. With the widespread adoption of internet electronic map services and location-based services (LBS) applications, POIs have undergone significant development in both conceptual scope and information depth.

[0003] Next point of interest (NPI) recommendations suggest destinations for users to visit in the next moment. NPI recommendation plays a crucial role in many location-based applications, providing users with personalized suggestions about attractive destinations. With the accumulation of massive amounts of mobile data, NPI recommendation has become an important task in location-based social networks. Summary of the Invention

[0004] According to a first aspect of this disclosure, an interest point recommendation method is provided, comprising: dividing the historical trajectories of a target user and other users into sessions according to a specified time interval, wherein a session includes multiple interest points; generating a global interaction hypergraph based on the sessions, wherein the global interaction hypergraph includes nodes and hyperedges, where nodes represent interest points and hyperedges represent connections between interest points in the same session; generating global interest point information based on the global interaction hypergraph; and generating target interest points to recommend to the target user based on the target user's initial information and the global interest point information.

[0005] In some embodiments, generating global interest point information based on a global interaction hypergraph includes: generating global interest point information using a hypergraph convolutional network based on the global interaction hypergraph.

[0006] In some embodiments, the interest point recommendation method, wherein generating global interest point information using a hypergraph convolutional network based on a global interaction hypergraph includes: generating hyperedge aggregation information using a hypergraph convolutional network based on an association matrix describing the connection relationships between nodes and hyperedges in the global interaction hypergraph; generating node aggregation information using a hypergraph convolutional network based on the hyperedge aggregation information and the transpose of the association matrix; and generating global interest point information based on the node aggregation information.

[0007] In some embodiments, aggregation information of hyperedges is generated using a hypergraph convolutional network based on an association matrix describing the connection relationships between nodes and hyperedges in a global interactive hypergraph. This includes: for the first layer of the hypergraph convolutional network, generating aggregation information of hyperedges in the first layer of the hypergraph convolutional network based on the association matrix and initial information of interest points; and for other layers of the hypergraph convolutional network, generating aggregation information of hyperedges in the current layer of the hypergraph convolutional network based on the association matrix and aggregation information of nodes in the previous layer.

[0008] In some embodiments, generating global interest point information based on node aggregation information includes: generating global interest point information based on the aggregation information of nodes in multiple layers of a hypergraph convolutional network.

[0009] In some embodiments, generating target points of interest recommended to the target user based on the target user's initial information and global point of interest information includes: generating spatiotemporal enhancement information of the points of interest based on the global point of interest information and the time information of the points of interest, wherein the time information of the points of interest includes information on the order in which the target user accesses the points of interest; generating spatiotemporal enhancement information of the target user based on the initial information of the points of interest in the target user's historical trajectory; and generating target points of interest recommended to the target user based on the spatiotemporal enhancement information of the points of interest and the spatiotemporal enhancement information of the target user.

[0010] In some embodiments, generating spatiotemporal augmentation information for a target user based on initial information of points of interest in the target user's historical trajectory includes: generating spatiotemporal augmentation information for the target user based on at least one of initial information, time information, and geographic location information of points of interest in the target user's historical trajectory, wherein the geographic location information of points of interest in the target user's historical trajectory includes distance constraints between points of interest.

[0011] In some embodiments, the point-of-interest recommendation method further includes: calculating the distance between the geographical locations of points of interest in the historical trajectory of the target user; constructing a geographic adjacency matrix based on the distance between the geographical locations of the points of interest; and generating the geographical location information of the points of interest in the historical trajectory of the target user based on the initial information of the points of interest in the historical trajectory of the target user and the geographic adjacency matrix.

[0012] In some embodiments, the target user's historical trajectory includes a time series of points of interest visited in the target user's history. The point of interest recommendation method further includes generating time information of the point of interest based on the initial information of the point of interest and the position encoding of the point of interest in the time series of points of interest visited in the target user's history.

[0013] In some embodiments, generating spatiotemporal augmentation information for a target user based on at least one of initial information, time information, and geographic location information of points of interest in the target user's historical trajectory includes: generating spatiotemporal augmentation information for the target user using a self-attention network based on a weighted sum of the initial information, time information, and geographic location information of points of interest in the target user's historical trajectory.

[0014] In some embodiments, generating spatiotemporal augmentation information of points of interest based on global point of interest information and time information of points of interest includes: generating local point of interest information based on initial information of the target user and initial information of points of interest in the target user's historical trajectory; generating spatial augmentation information of points of interest based on local point of interest information and global point of interest information; and generating spatiotemporal augmentation information of points of interest based on spatial augmentation information of points of interest and time information of points of interest.

[0015] According to a second aspect of this disclosure, a training method for an interest point recommendation model is provided, comprising: using the interest point recommendation model, dividing the historical trajectories of a target user and other users into sessions according to specified time intervals, wherein a session includes multiple interest points; generating a global interaction hypergraph based on the sessions, wherein the global interaction hypergraph includes nodes and hyperedges, nodes representing interest points and hyperedges representing connections between interest points in the same session; generating global interest point information based on the global interaction hypergraph; generating a prediction result of target interest points to be recommended to the target user based on the initial information of the target user and the global interest point information; and training the interest point recommendation model based on the prediction result of the target interest points to be recommended to the target user and the true values ​​of the target interest points.

[0016] According to a third aspect of this disclosure, an interest point recommendation device is provided, comprising: a trajectory processing module configured to divide the historical trajectories of a target user and other users into sessions according to a specified time interval, wherein a session includes multiple interest points; a hypergraph generation module configured to generate a global interactive hypergraph based on the sessions, wherein the global interactive hypergraph includes nodes and hyperedges, where nodes represent interest points and hyperedges represent connections between interest points in the same session; a global interest point information generation module configured to generate global interest point information based on the global interactive hypergraph; and a target interest point generation module configured to generate target interest points to be recommended to the target user based on the target user's initial information and the global interest point information.

[0017] According to a fourth aspect of this disclosure, a training apparatus for an interest point recommendation model is provided, comprising: a trajectory processing module configured to use the interest point recommendation model to divide the historical trajectories of a target user and other users into sessions at specified time intervals, wherein a session includes multiple interest points; a hypergraph generation module configured to generate a global interaction hypergraph based on the sessions, wherein the global interaction hypergraph includes nodes and hyperedges, where nodes represent interest points and hyperedges represent connections between interest points in the same session; a global interest point information generation module configured to generate global interest point information based on the global interaction hypergraph; a target interest point generation module configured to generate a prediction result of target interest points recommended to the target user based on the initial information of the target user and the global interest point information; and a training module configured to train the interest point recommendation model based on the prediction result of the target interest points recommended to the target user and the true values ​​of the target interest points.

[0018] According to a fifth aspect of this disclosure, an electronic device is provided, comprising: a memory; and a processor coupled to the memory, the processor being configured to execute an interest point recommendation method or an interest point recommendation model training method according to any embodiment of this disclosure based on instructions stored in the memory.

[0019] According to a sixth aspect of this disclosure, a computer-readable storage medium is provided that stores computer program instructions thereon, which, when executed by a processor, implement the interest point recommendation method or the interest point recommendation model training method according to any embodiment of this disclosure.

[0020] According to a seventh aspect of this disclosure, a spatiotemporal modeling system is provided, including an interest point recommendation method or an interest point recommendation model training method according to any embodiment of this disclosure. Attached Figure Description

[0021] The accompanying drawings, which form part of this specification, illustrate embodiments of this disclosure and, together with the specification, serve to explain the principles of this disclosure.

[0022] This disclosure will become clearer with reference to the accompanying drawings and the following detailed description, wherein:

[0023] Figure 1 A flowchart illustrating a point-of-interest recommendation method according to some embodiments of the present disclosure is shown;

[0024] Figure 2 A schematic diagram illustrating user history trajectories according to some embodiments of the present disclosure;

[0025] Figure 3 A schematic diagram illustrating the historical trajectory division according to some embodiments of this disclosure is shown;

[0026] Figure 4 A schematic diagram of a hypergraph convolutional network according to some embodiments of the present disclosure is shown;

[0027] Figure 5 A schematic diagram of a local spatiotemporal convolutional network according to some embodiments of the present disclosure is shown;

[0028] Figure 6 A schematic diagram illustrating target interest point prediction according to some embodiments of the present disclosure is shown;

[0029] Figure 7 A flowchart illustrating the training of an interest point recommendation model according to some embodiments of the present disclosure is shown;

[0030] Figure 8 A block diagram of a point-of-interest recommendation apparatus according to some embodiments of the present disclosure is shown;

[0031] Figure 9 A block diagram of a point-of-interest recommendation apparatus according to some embodiments of the present disclosure is shown;

[0032] Figure 10 Block diagrams of electronic devices according to other embodiments of the present disclosure are shown;

[0033] Figure 11 A block diagram of a computer system for implementing some embodiments of the present disclosure is shown. Detailed Implementation

[0034] Various exemplary embodiments of the present disclosure will now be described in detail with reference to the accompanying drawings. It should be noted that, unless otherwise specifically stated, the relative arrangement, numerical expressions, and values ​​of the components and steps set forth in these embodiments do not limit the scope of the present disclosure.

[0035] At the same time, it should be understood that, for ease of description, the dimensions of the various parts shown in the accompanying drawings are not drawn according to actual scale.

[0036] The following description of at least one exemplary embodiment is merely illustrative and is in no way intended to limit this disclosure or its application or use.

[0037] Techniques, methods, and equipment known to those skilled in the art may not be discussed in detail, but where appropriate, such techniques, methods, and equipment should be considered part of the specification.

[0038] In all examples shown and discussed herein, any specific values ​​should be interpreted as merely exemplary and not as limitations. Therefore, other examples of exemplary embodiments may have different values.

[0039] It should be noted that similar labels and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be discussed further in subsequent figures.

[0040] In related technologies, in order to capture cooperative signals, one-hop neighbors of interest points are randomly sampled, but higher-order signals between interest points are ignored, resulting in data sparsity problems.

[0041] Furthermore, related technologies neglect spatiotemporal relationships when recommending points of interest. When users access the same points of interest in different sequences, simply aggregating the interacting points results in identical learned embeddings. However, since users are in different sequences, their representations should differ, indicating a lack of understanding of spatiotemporal relationships.

[0042] Figure 1 A flowchart illustrating a point-of-interest recommendation method according to some embodiments of this disclosure is shown.

[0043] like Figure 1 As shown, the point of interest recommendation method includes steps S11-S14.

[0044] In step S11, the historical trajectories of the target user and other users are divided into sessions according to a specified time interval, wherein a session includes multiple points of interest.

[0045] In traditional graph structures, each edge can only connect two vertices. However, in hypergraphs, a new type of edge is abstracted: the hyperedge. A hyperedge can connect more than two vertices simultaneously. Hyperedges can have duplicate nodes. Therefore, hypergraphs transcend pairwise node relationships to unify nodes.

[0046] Figure 2 A schematic diagram illustrating user history trajectories according to some embodiments of this disclosure is shown.

[0047] like Figure 2 As shown, each user's historical trajectory sequence is divided into multiple sessions according to a specified time interval. For example, a session represents a set of points of interest that a user visited in a day.

[0048] Figure 3 A schematic diagram illustrating the historical trajectory division according to some embodiments of this disclosure is shown.

[0049] like Figure 3 As shown, in user 2's trajectory graph, l1 and l4 belong to different sessions and have a multi-hop relationship. By constructing a hypergraph by dividing sessions according to time intervals, the multi-hop relationship between points of interest can be captured.

[0050] In step S12, a global interactive hypergraph is generated based on the session. The global interactive hypergraph includes nodes and hyperedges. Nodes represent points of interest, and hyperedges represent connections between points of interest in the same session.

[0051] Figure 2 In the trajectory of user 1, restaurants and stadiums are high-order neighbors of the coffee shop. For user 4, the high-order neighbors of the coffee shop are cinemas and shopping malls. Therefore, restaurants, stadiums, cinemas, and shopping malls are potentially related, and there may be implicit high-order collaborative information between them. By constructing a hypergraph to capture such information, the data sparsity problem can be alleviated.

[0052] Figure 4 A schematic diagram of a hypergraph convolutional network according to some embodiments of the present disclosure is shown.

[0053] like Figure 4 As shown, the interest point recommendation model includes a hypergraph convolutional network, also known as a global interaction hypergraph neural network. Figure 4 In this context, "embedded" refers to "information," for example, global POI embedding represents global point of interest information. Figure 5 and Figure 6 Similarly. In Figure 4 In the first step, an interactive hypergraph is constructed. To reveal higher-order coordination signals between sessions. In the hypergraph In the middle, points of interest are represented by nodes l i This indicates that points of interest nodes in a session are connected together using hyperedges.

[0054] In step S13, global interest point information (i.e., global interest point embedding) is generated based on the global interaction hypergraph.

[0055] For example, after constructing a global interaction hypergraph, the hypergraph network is used to capture the relationships between high-order interest points in the global interaction hypergraph in an iterative manner.

[0056] In some embodiments, generating global interest point information based on a global interaction hypergraph includes: generating global interest point information using a hypergraph convolutional network based on the global interaction hypergraph.

[0057] In some embodiments, generating global interest point information using a hypergraph convolutional network based on a global interaction hypergraph includes: generating hyperedge aggregation information using a hypergraph convolutional network based on an association matrix describing the connection relationships between nodes and hyperedges in the global interaction hypergraph; generating node aggregation information using a hypergraph convolutional network based on the hyperedge aggregation information and the transpose of the association matrix; and generating global interest point information based on the node aggregation information.

[0058] For example, introducing an association matrix This describes the connections between nodes and hyperedges, thus constructing the topological structure of the hypergraph. An element in the incidence matrix H is defined as follows:

[0059]

[0060] in, It is a node, e∈ε H It is a hyperedge, and all h(v,e) form the incidence matrix H.

[0061] For each edge e, the degree of the edge is denoted as d(e), and the calculation formula is as follows.

[0062]

[0063] For each node v, the degree of the node is represented by d(v), and the calculation formula is as follows:

[0064]

[0065] The degrees of all nodes form the diagonal degree matrix D. H The degrees of all hyperedges form the diagonal hyperedge degree matrix B.

[0066] Using a two-step propagation method, the hyperedge acts as a medium for the aggregation of nodes within the hyperedge and the propagation across the hyperedge, aggregating information about nodes and hyperedges in the hypergraph.

[0067] The first step is to aggregate node representations within each hyperedge. H represents the polymerization process. T It reflects the relationships between superedge nodes.

[0068] The second step, after aggregating node representations within each hyperedge, involves pre-multiplying by H to aggregate information from the hyperedge to the nodes. Since the association matrix H represents the node-hyperedge relationship, the hyperedge-to-node propagation phase aims to enrich the representation of node information by utilizing global information outside the current hyperedge.

[0069] In some embodiments, the aggregation information of hyperedges is generated using a hypergraph convolutional network based on the association matrix, including: for the first layer of the hypergraph convolutional network, generating the aggregation information of hyperedges in the first layer of the hypergraph convolutional network based on the association matrix and the initial information of interest points; for other layers of the hypergraph convolutional network, generating the aggregation information of hyperedges in the current layer of the hypergraph convolutional network based on the association matrix and the aggregation information of nodes in the previous layer of the hypergraph convolutional network.

[0070] In some embodiments, generating global interest point information based on the aggregation information of multiple nodes includes: generating global interest point information based on the aggregation information of multiple nodes in multiple layers of a hypergraph convolutional network.

[0071] For example, given a target user u and its trajectory sequence S u The goal of the next point of interest recommendation is to recommend the top K points of interest that u is likely to visit at the next timestamp.

[0072] Sort user u's interactive points of interest according to the time of their access to those points. The set of points of interest corresponding to user u after sorting is represented as follows: Where m is the sequence length of user u. Then, by looking up the table, we obtain... Initial information of the corresponding points of interest (initial embedding) in Let E represent the information dimension as d. u It serves as the input to the first layer of a hypergraph convolutional network.

[0073] The hypergraph convolutional network is represented as follows:

[0074]

[0075] in, This represents the information of all interest points encoded by the (k+1)th hypergraph convolutional network layer. W is a diagonal weight matrix with positive weights, and B is a matrix of diagonal hypergraph edges. D H A matrix with diagonal nodes.

[0076] In some embodiments, generating global interest point information based on the aggregation information of multiple nodes includes: generating global interest point information based on the aggregation information of multiple nodes in multiple layers of a hypergraph convolutional network.

[0077] For example, the outputs of multiple layers of a hypergraph convolutional network are averaged using pooling and other methods, and the averaged result is used as global interest point information X. H .

[0078] In step S14, target interest points recommended to the target user are generated based on the target user's initial information and global interest point information.

[0079] The initial information of the target user is the initial embedding representation of the target user. Based on the initial information of the target user and the previously calculated global interest point information, the next target interest point is predicted.

[0080] In some embodiments, generating target points of interest recommended to the target user based on the target user's initial information and global point of interest information includes: generating spatiotemporal enhancement information of the points of interest based on the global point of interest information and the time information of the points of interest, wherein the time information of the points of interest includes information on the order in which the target user accesses the points of interest; generating spatiotemporal enhancement information of the target user based on the initial information of the points of interest in the target user's historical trajectory; and generating target points of interest recommended to the target user based on the spatiotemporal enhancement information of the points of interest and the spatiotemporal enhancement information of the target user.

[0081] For example, User 1 and User 2 visit the same points of interest (coffee shop, office, restaurant, and stadium), but in a different order. If we simply aggregate the interacting points of interest, the embeddings of these two users would be identical. However, due to the different order, the latent representations of User 1 and User 2 should be different. By incorporating the spatiotemporally augmented information of the target user (i.e., spatiotemporally augmented embeddings), the point of interest recommendation model can learn the user's preferences for the prediction.

[0082] Therefore, in addition to the hypergraph, a local spatiotemporal augmentation graph and a local spatiotemporal convolutional network (also known as a spatiotemporal neural network with local spatiotemporal enhancement) are constructed. When generating spatiotemporal augmentation information for the target user, time information and / or spatial information are added to aggregate and propagate user and point of interest nodes in an asymmetric manner to capture the spatiotemporal augmentation dependencies between the target user's points of interest, including complex user-point of interest interactions, point of interest-point of interest sequence relationships, and non-adjacent point of interest-point of interest geographical relationships.

[0083] Figure 5 A schematic diagram of a local spatiotemporal convolutional network according to some embodiments of the present disclosure is shown.

[0084] like Figure 5 As shown, the interest point recommendation model also includes a local spatiotemporal convolutional network. First, a local spatiotemporal augmentation map is constructed. Where node v L Including user node u i and point of interest nodes l i , edge ε L It consists of user-point-of-interest (POI) interactions, POI-POI temporal relationships, and non-adjacent POI-POI geographical relationships.

[0085] In some embodiments, the target user's historical trajectory includes a time series of points of interest visited in the target user's history. The point of interest recommendation method further includes generating time information of the point of interest based on the position encoding of the point of interest in the time series of points of interest visited in the target user's history and the initial information of the point of interest.

[0086] For example, multiple points of interest (POIs) in an interaction may have temporal dependencies. Adding temporal information to the initial information of POIs yields temporal information (i.e., temporal embedding). For instance, using position encoding in attention mechanisms, temporal information is equivalent to position embedding within the attention mechanism. This temporal information is represented as the positional information of a POI within the user's access trajectory sequence, thus representing the sequential relationship between POIs and achieving sequence modeling. The temporal information of the set of POIs after incorporating temporal information is denoted as... in,

[0087] In some embodiments, the point-of-interest recommendation method further includes: calculating the distance between the geographical locations of points of interest in the historical trajectory of the target user; constructing a geographic adjacency matrix based on the distance between the geographical locations of the points of interest; and generating the geographical location information of the points of interest in the historical trajectory of the target user based on the initial information of the points of interest in the historical trajectory of the target user and the geographic adjacency matrix.

[0088] For example, in Figure 2 In this example, the distance between two adjacent points of interest for user 3 is 25km, meaning that user 3's spatial acceptability is at least 25km. By generating geographic location information (i.e., geographic location embedding), the user's spatial acceptability can be taken into account.

[0089] When generating geographic information, construct a geographic adjacency matrix. To reflect the boundary constraints between interactive points of interest. Define the geographic adjacency matrix A. geo In the diagram, the element at position (I, j) is defined as geographic influence a. ij The calculation method is as follows.

[0090] a ij =exp(-dist(d i ,d j ) 2 )

[0091] in, It is set T u A set of interest point pairs, d i ,d j They are respectively The geographic coordinates (e.g., longitude and latitude). `dist` represents the hadrsine distance. The hadrsine equation is used to calculate the distance between two points at different latitudes and longitudes, and it is expressed as follows:

[0092]

[0093] Where r is the radius, lat2 and lat1 represent longitude, and lon2 and lon1 represent latitude.

[0094] A Gaussian kernel function is used to represent the geographic influence between two points of interest, and to describe the negative correlation between geographic influence and distance, with the constraint range controlled between 0 and 1. Furthermore, a graph convolutional network is used to capture the nonlinear geographic influence between interacting points of interest; the calculation formula is shown below.

[0095] V u =A geo E u W geo +b geo

[0096] Among them, V u Representing geographic information, W geo Let b represent the transition matrix. geo It is the bias vector. V u This represents the non-linear geographic influence between points of interest.

[0097] In some embodiments, generating spatiotemporal augmentation information for a target user based on at least one of initial information, time information, and geographic location information of points of interest in the target user's historical trajectory includes: generating spatiotemporal augmentation information for the target user using a self-attention network based on a weighted sum of the initial information, time information, and geographic location information of points of interest in the target user's historical trajectory.

[0098] For example, local spatiotemporal convolutional networks include self-attention networks. Element-wise addition is performed on the target user's initial information, temporal information, and geographic information to obtain the target user's spatiotemporal information (i.e., spatiotemporal embedding) Z. u =E u +P u +V u .

[0099] When generating spatiotemporal augmentation information for target users, time information is incorporated to take into account the time sequence of target users' access to points of interest, thereby modeling the spatiotemporal correlation between points of interest and making the recommendation results more accurate.

[0100] When generating spatiotemporal augmentation information for target users, geographic location information is incorporated, taking into account the spatial acceptability of target users and increasing the spatial influence between points of interest. This models the spatiotemporal correlation between points of interest, making the recommendation results more accurate.

[0101] Then, the self-attention mechanism in sequence modeling methods is used to capture the interdependencies of interest points in the interest point sequence. For example, Z... uAs input, the data is processed through a multi-head self-attention network of a local spatiotemporal convolutional network and a feedforward neural network (FNN) to obtain the spatiotemporal augmentation information of the target user (i.e., spatiotemporal augmentation embedding).

[0102]

[0103] Figure 6 A schematic diagram illustrating target interest point prediction according to some embodiments of the present disclosure is shown.

[0104] like Figure 6 As shown, spatiotemporal augmentation information of the points of interest is generated based on global point of interest information and the temporal information of the points of interest, including: generating local point of interest information based on the initial information of the target user and the initial information of the points of interest in the target user's historical trajectory; generating spatial augmentation information of the points of interest based on the local point of interest information and global point of interest information; and generating spatiotemporal augmentation information of the points of interest based on the spatial augmentation information of the points of interest and the temporal information of the points of interest.

[0105] For example, based on initial user information (initial user embedding) and initial point of interest information (initial point of interest embedding), local point of interest information is generated through a neighborhood aggregation network.

[0106] Local point of interest information and global interest point information X H By adding elements one by one, we obtain the spatial augmentation information of the interest points (i.e., the interest point embedding) X. L The formula is shown below.

[0107]

[0108] Then, based on the spatial augmentation information and the reverse temporal information (i.e., reverse position embedding), spatiotemporal augmentation information (i.e., spatiotemporal augmentation embedding) is generated. The formula for representing the spatiotemporal enhancement information of the i-th interest point in the interest point ranking sequence of user u is as follows.

[0109]

[0110] in, For X L The index. Indicates will and p m+1-i The concatenation, W1, and b1 are trainable parameters of the interest point prediction model. m+1-i Representing reversed time information is time information P. uThe elements in the model are arranged from last to first according to the time when the target user visited the corresponding point of interest. By adding reverse time information, the point of interest prediction model pays more attention to the places the target user has recently visited, thus improving the accuracy of the prediction.

[0111] By providing By assigning different attention weights, incremental information (i.e., incremental embedding) of the target user at time T (i.e., the target time) is obtained. The formula is shown below.

[0112]

[0113]

[0114] Where q, W2, W3, and b2 are trainable attention parameters. σ represents the activation function. It is based on all global interest point information X of user u H It is aggregated using mean pooling.

[0115] In acquiring spatiotemporal augmentation information of target user u and incremental information Then, the two are added together to calculate the final user representation u, as shown in the following formula.

[0116]

[0117] Then, using a predictor, based on the final user representation u and the target interest point representation x... l Generate prediction results The calculation formula is shown below.

[0118]

[0119] According to some embodiments of the interest point recommendation method disclosed herein, the historical trajectories of a target user and other users are divided into sessions at specified time intervals, wherein a session includes multiple interest points; a global interaction hypergraph is generated based on the sessions, wherein the global interaction hypergraph includes nodes and hyperedges, where nodes represent interest points and hyperedges represent connections between interest points within the same session; global interest point information is generated based on the global interaction hypergraph; and target interest points recommended to the target user are generated based on the target user's initial information and the global interest point information. By constructing a hypergraph and calculating global interest point information, higher-order signals between interest points in different sessions can be captured, alleviating the problem of data sparsity and improving the accuracy of interest point recommendation. Furthermore, by constructing the hypergraph by dividing the historical trajectory into sessions, multi-hop relationships between interest points can be captured, further improving the accuracy of interest point recommendation.

[0120] Figure 7 A flowchart illustrating the training of an interest point recommendation model according to some embodiments of this disclosure is shown.

[0121] like Figure 7 As shown, the training method for the interest point recommendation model includes steps S21-S25.

[0122] In step S21, using the point of interest recommendation model, the historical trajectories of the target user and other users are divided into sessions according to a specified time interval.

[0123] In step S22, a global interactive hypergraph is generated based on the session. The global interactive hypergraph includes nodes and hyperedges. Nodes represent points of interest, and hyperedges represent connections between points of interest in the same session.

[0124] In step S23, global interest point information is generated based on the global interactive hypergraph.

[0125] In step S24, based on the initial information of the target user and the global interest point information, a prediction result of the target interest points recommended to the target user is generated.

[0126] In step S25, an interest point recommendation model is trained based on the predicted results of the target interest points recommended to the target user and the actual values ​​of the target interest points.

[0127] Steps S21-S24 are similar to steps S11-S14, and will not be repeated here to avoid repetition.

[0128] In step S25, based on the predicted results of the target interest points recommended to the target user and the actual values ​​of the target interest points, a loss function is calculated to train the interest point recommendation model. The loss function, for example, is the cross-entropy loss function, and its calculation formula is shown below.

[0129]

[0130] in, It is an indicator function representing 0 and 1, indicating when the user visits the next point of interest. j hour, Equals 1 if true, otherwise 0 if false. ||Θ||2 represents L2 regularization. λ||Θ||2 means that under the control of λ, L2 regularization is used on all parameters to prevent overfitting.

[0131] According to the training method of the interest point recommendation model according to some embodiments of this disclosure, the historical trajectories of the target user and other users are divided into sessions according to a specified time interval, wherein a session includes multiple interest points; a global interaction hypergraph is generated based on the sessions, wherein the global interaction hypergraph includes nodes and hyperedges, nodes represent interest points, and hyperedges represent connections between interest points in the same session; global interest point information is generated based on the global interaction hypergraph; a prediction result of the target interest points recommended to the target user is generated based on the initial information of the target user and the global interest point information; and the interest point recommendation model is trained based on the prediction result of the target interest points recommended to the target user and the true value of the target interest points. By constructing a hypergraph and calculating global interest point information, high-order signals between interest points in different sessions can be captured, alleviating the problem of data sparsity and improving the prediction accuracy of the interest point recommendation model. In addition, by dividing the historical trajectory into sessions to construct the hypergraph, multi-hop relationships between interest points can be captured, further improving the prediction accuracy of the interest point recommendation model.

[0132] Figure 8 A block diagram of a point-of-interest recommendation apparatus according to some embodiments of the present disclosure is shown.

[0133] like Figure 8 As shown, the point of interest recommendation device 8 includes a trajectory processing module 81, a hypermap generation module 82, a global point of interest information generation module, and a target point of interest generation module 84.

[0134] The trajectory processing module 81 is configured to divide the historical trajectories of the target user and other users into sessions according to a specified time interval, for example, by performing the following... Figure 1 Step S11 is shown.

[0135] Hypergraph generation module 82 is configured to generate a global interactive hypergraph based on a session. The global interactive hypergraph includes nodes and hyperedges; nodes represent points of interest, and hyperedges represent connections between points of interest within the same session. For example, when performing an action... Figure 1 Step S12 is shown.

[0136] The global point of interest (POI) information generation module 83 is configured to generate global point of interest information based on the global interactive hypergraph, for example, by executing... Figure 1 Step S13 is shown.

[0137] The target interest point generation module 84 is configured to generate target interest points recommended to the target user based on the target user's initial information and global interest point information, for example, by executing... Figure 1 Step S14 is shown.

[0138] In some embodiments, the global interest point information generation module 83 is further configured to: generate global interest point information using a hypergraph convolutional network based on the global interaction hypergraph.

[0139] In some embodiments, the global interest point information generation module 83 is further configured to: generate hyperedge aggregation information using a hypergraph convolutional network based on an association matrix describing the connection relationships between nodes and hyperedges in the global interactive hypergraph; generate node aggregation information using a hypergraph convolutional network based on the hyperedge aggregation information and the transpose of the association matrix; and generate global interest point information based on the node aggregation information.

[0140] In some embodiments, the global interest point information generation module 83 is further configured to: for the first layer of the hypergraph convolutional network, generate aggregation information of hyperedges in the first layer of the hypergraph convolutional network based on the association matrix and the initial information of the interest points; for other layers of the hypergraph convolutional network besides the first layer, generate aggregation information of hyperedges in the current layer of the hypergraph convolutional network based on the association matrix and the aggregation information of nodes in the previous layer.

[0141] In some embodiments, the global interest point information generation module 83 is further configured to generate global interest point information based on the aggregation information of nodes in the multilayer of the hypergraph convolutional network.

[0142] In some embodiments, the target point of interest generation module 84 is further configured to: generate spatiotemporal enhancement information of the point of interest based on global point of interest information and the time information of the point of interest, wherein the time information of the point of interest includes information on the order in which the target user accesses the point of interest; generate spatiotemporal enhancement information of the target user based on the initial information of the point of interest in the target user's historical trajectory; and generate target points of interest recommended to the target user based on the spatiotemporal enhancement information of the point of interest and the spatiotemporal enhancement information of the target user.

[0143] In some embodiments, the target point of interest generation module 84 is further configured to generate spatiotemporal augmentation information of the target user based on at least one of the initial information, time information, and geographic location information of the points of interest in the target user's historical trajectory, wherein the geographic location information of the points of interest in the target user's historical trajectory includes distance constraints between the points of interest.

[0144] In some embodiments, the point-of-interest recommendation model further includes: a geographic location information generation module, configured to calculate the distance between geographic locations of points of interest in the historical trajectory of a target user; construct a geographic adjacency matrix based on the distance between geographic locations of points of interest; and generate geographic location information of points of interest in the historical trajectory of a target user based on the initial information of points of interest in the historical trajectory of the target user and the geographic adjacency matrix.

[0145] In some embodiments, the target user's historical trajectory includes a time series of points of interest (POIs) visited historically by the target user. The POI recommendation model further includes a time information generation module configured to generate time information of the POIs based on the initial information of the POIs and the positional encoding of the POIs in the time series of POIs visited historically by the target user. In some embodiments, the target POI generation module 84 is further configured to generate spatiotemporal augmentation information of the target user based on at least one of the initial information, time information, and geographic location information of the POIs in the target user's historical trajectory, including generating the spatiotemporal augmentation information of the target user using a self-attention network based on a weighted sum of the initial information, time information, and geographic location information of the POIs in the target user's historical trajectory.

[0146] The target point of interest generation module 84 is further configured to: generate local point of interest information based on the initial information of the target user and the initial information of the points of interest in the target user's historical trajectory; generate spatial enhancement information of the points of interest based on the local point of interest information and the global point of interest information; and generate spatiotemporal enhancement information of the points of interest based on the spatial enhancement information of the points of interest and the temporal information of the points of interest.

[0147] According to some embodiments of the present disclosure, the point-of-interest (POI) recommendation apparatus, by constructing a hypergraph and calculating global POI information, can capture high-order signals between POIs in different sessions, alleviating the problem of data sparsity and improving the accuracy of POI recommendation. Furthermore, by constructing a hypergraph by dividing historical trajectories into sessions, it can capture multi-hop relationships between POIs, further improving the accuracy of POI recommendation.

[0148] Figure 9 A block diagram of a point-of-interest recommendation apparatus according to some embodiments of the present disclosure is shown.

[0149] like Figure 9 As shown, the point of interest recommendation device 9 includes a trajectory processing module 91, a hypergraph generation module 92, a global point of interest information generation module 93, a target point of interest generation module 94, and a training module 95.

[0150] The trajectory processing module 91 is configured to utilize an interest point recommendation model to divide the historical trajectories of the target user and other users into sessions according to specified time intervals, for example, by performing... Figure 7 Step S21 is shown.

[0151] Hypergraph generation module 92 is configured to generate a global interactive hypergraph based on a session. The global interactive hypergraph includes nodes and hyperedges; nodes represent points of interest, and hyperedges represent connections between points of interest within the same session. For example, when performing an action... Figure 7 Step S22 is shown.

[0152] The global point of interest (POI) information generation module 93 is configured to generate global point of interest information based on the global interactive hypergraph, for example, by executing... Figure 7 Step S23 is shown.

[0153] The target interest point generation module 94 is configured to generate a prediction result of target interest points recommended to the target user based on the target user's initial information and global interest point information, for example, by performing the following... Figure 7 Step S24 is shown.

[0154] Training module 95 is configured to train an interest point recommendation model based on the predicted values ​​of target interest points recommended to the target user and the actual values ​​of the target interest points, for example, by performing actions such as... Figure 7 Step S25 is shown.

[0155] The trajectory processing module 91, hypermap generation module 92, global interest point information generation module 93, and target interest point generation module 94 are similar to the trajectory processing module 81, hypermap generation module 82, global interest point information generation module 83, and target interest point generation module 84, and will not be described in detail here.

[0156] The training apparatus for the point of interest (POI) recommendation model according to some embodiments of this disclosure, by constructing a hypergraph and calculating global POI information, can capture high-order signals between POIs in different sessions, alleviating the problem of data sparsity and improving the prediction accuracy of the POI recommendation model. Furthermore, by constructing a hypergraph by dividing historical trajectories into sessions, it is possible to capture multi-hop relationships between POIs, further improving the prediction accuracy of the POI recommendation model.

[0157] Figure 10 Block diagrams of electronic devices according to other embodiments of the present disclosure are shown.

[0158] like Figure 10 As shown, the electronic device 10 includes a memory 101 and a processor 102 coupled to the memory 101. The memory 101 stores training methods for executing point-of-interest (POI) recommendation methods or POI recommendation models. The processor 102 is configured to execute POI recommendation methods or POI recommendation model training methods in any of the embodiments of this disclosure based on instructions stored in the memory 101.

[0159] Figure 11 A block diagram of a computer system for implementing some embodiments of the present disclosure is shown.

[0160] like Figure 11 As shown, the computer system 110 can be represented in the form of a general computing device. The computer system 110 includes a memory 1111, a processor 1120, and a bus 1100 connecting different system components.

[0161] The memory 1111 may include, for example, system memory, non-volatile storage media, etc. System memory may store, for example, an operating system, application programs, a boot loader, and other programs. System memory may include volatile storage media, such as random access memory (RAM) and / or cache memory. Non-volatile storage media may store, for example, instructions for executing the interest point recommendation method or the training method of the interest point recommendation model in any of the embodiments of this disclosure. Non-volatile storage media include, but are not limited to, disk storage, optical storage, flash memory, etc.

[0162] The processor 1120 can be implemented using a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA) or other programmable logic devices, discrete hardware components such as discrete gates or transistors. Accordingly, each module, such as the decision module and the determination module, can be implemented by executing instructions in the central processing unit (CPU) memory to perform the corresponding steps, or by implementing dedicated circuitry to perform the corresponding steps.

[0163] Bus 1100 can use any of the various bus architectures. For example, bus architectures include, but are not limited to, Industry Standard Architecture (ISA) bus, Micro Channel Architecture (MCA) bus, and Peripheral Component Interconnect (PCI) bus.

[0164] The computer system 110 may also include an input / output interface 1130, a network interface 1140, and a storage interface 1150. These interfaces 1130, 1140, and 1150, as well as the memory 1111 and processor 1120, can be connected via a bus 1100. The input / output interface 1130 provides a connection interface for input / output devices such as a monitor, mouse, and keyboard. The network interface 1140 provides a connection interface for various networked devices. The storage interface 1150 provides a connection interface for external storage devices such as floppy disks, USB flash drives, and SD cards.

[0165] This disclosure also provides a spatiotemporal modeling system, including an interest point recommendation device or an interest point recommendation model training device according to any embodiment of this disclosure. The spatiotemporal modeling system according to this disclosure improves the accuracy of interest point recommendation by constructing a hypergraph to complete spatiotemporal modeling.

[0166] Various aspects of this disclosure are described herein with reference to flowchart illustrations and / or block diagrams of methods, apparatus, and computer program products according to embodiments of this disclosure. It should be understood that each block of the flowchart illustrations and / or block diagrams, and combinations thereof, can be implemented by computer-readable program instructions.

[0167] These computer-readable program instructions are provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable device to produce a machine, such that execution of the instructions by the processor produces means for implementing the functions specified in one or more boxes of the flowchart and / or block diagram.

[0168] These computer-readable program instructions are also readablely stored in a computer-readable storage medium. These instructions cause a computer to work in a particular manner to produce an article of manufacture, including instructions that implement the functions specified in one or more boxes in a flowchart and / or block diagram.

[0169] This disclosure may take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects.

[0170] The interest point recommendation method or interest point recommendation model training method and apparatus, computer-readable storage medium, and spatiotemporal modeling system in the above embodiments improve the accuracy of interest point recommendation.

[0171] This concludes the detailed description of the point-of-interest recommendation method and apparatus, computer-readable storage medium, and spatiotemporal modeling system according to this disclosure. To avoid obscuring the concept of this disclosure, some details known in the art have not been described. Those skilled in the art will fully understand how to implement the technical solutions disclosed herein based on the above description.

Claims

1. An interest-based recommendation method, comprising: The historical trajectories of the target user and other users are divided into sessions according to a specified time interval. Each session includes multiple points of interest. Based on the session, a global interaction hypergraph is generated, which includes nodes and hyperedges. Nodes represent points of interest, and hyperedges represent connections between points of interest in the same session. Generate global point of interest information based on the global interactive hypergraph; Based on the target user's initial information and global point of interest (POI) information, generate target POIs recommended to the target user, including: generating spatiotemporal augmentation information for the POIs based on the global POI information and the POIs' temporal information, wherein the POIs' temporal information includes the order in which the target user accessed the POIs; generating spatiotemporal augmentation information for the target user based on the initial information of the POIs in the target user's historical trajectory; and generating target POIs recommended to the target user based on the spatiotemporal augmentation information of the POIs and the target user's spatiotemporal augmentation information. 2.The point of interest recommendation method of claim 1, wherein, Based on the global interaction hypergraph, global point of interest information is generated, including: Based on the global interaction hypergraph, global interest point information is generated using a hypergraph convolutional network. 3.The point of interest recommendation method of claim 2, wherein, Based on the global interaction hypergraph, global interest point information is generated using a hypergraph convolutional network, including: Based on the association matrix describing the connection relationships between nodes and hyperedges in the global interactive hypergraph, hyperedge aggregation information is generated using a hypergraph convolutional network. Based on the aggregation information of hyperedges and the transpose of the correlation matrix, the aggregation information of nodes is generated using a hypergraph convolutional network. Global point of interest information is generated based on the aggregation information of the nodes. 4.The point of interest recommendation method of claim 3, wherein, Based on the association matrix describing the connection relationships between nodes and hyperedges in the global interaction hypergraph, hyperedge aggregation information is generated using a hypergraph convolutional network, including: For the first layer of the hypergraph convolutional network, the aggregation information of the hyperedges in the first layer of the hypergraph convolutional network is generated based on the initial information of the correlation matrix and interest points. For layers other than the first layer of the hypergraph convolutional network, the aggregation information of the hyperedges in the current layer of the hypergraph convolutional network is generated based on the association matrix and the aggregation information of the nodes in the previous layer.

5. The point-of-interest recommendation method according to claim 3, wherein, Based on the aggregation information of the nodes, global point of interest information is generated, including: Global interest point information is generated based on the aggregation information of nodes in multiple layers of the hypergraph convolutional network.

6. The point-of-interest recommendation method according to claim 1, wherein, The target user's historical trajectory includes a time series of the target user's historical visits to points of interest. Point-of-interest recommendation methods also include: The time information of the point of interest is generated based on the initial information of the point of interest and the position encoding of the point of interest in the time series of the points of interest visited in the target user's history.

7. The point-of-interest recommendation method according to claim 1, wherein, Based on the initial information of points of interest in the target user's historical trajectory, spatiotemporal augmentation information for the target user is generated, including: Based on at least one of the initial information, time information, and geographic location information of the target user's historical trajectory, spatiotemporal augmentation information is generated, wherein the geographic location information of the target user's historical trajectory includes distance constraints between the points of interest.

8. The point-of-interest recommendation method according to claim 7, further comprising: Calculate the distances between geographical locations of points of interest in the target user's historical trajectory; Construct a geographic adjacency matrix based on the distance between the geographical locations of points of interest; Based on the initial information of points of interest in the target user's historical trajectory and the geographic adjacency matrix, the geographic location information of the points of interest in the target user's historical trajectory is generated.

9. The point-of-interest recommendation method according to claim 7 or 8, wherein, Based on at least one of the initial information, time information, and geographic location information of the target user's historical trajectory, generate spatiotemporal augmentation information for the target user, including: Based on the weighted sum of initial, temporal, and geographic information of points of interest in the target user's historical trajectory, a self-attention network is used to generate spatiotemporal augmentation information for the target user.

10. The point-of-interest recommendation method according to any one of claims 1, 6 to 8, wherein, Based on global point of interest (POI) information and POI temporal information, spatiotemporal augmentation information for POIs is generated, including: Based on the initial information of the target user and the initial information of the points of interest in the target user's historical trajectory, local point of interest information is generated; Based on local and global interest point information, spatial augmentation information for interest points is generated. Based on the spatial and temporal information of the point of interest, spatiotemporal enhancement information of the point of interest is generated.

11. A training method for an interest point recommendation model, comprising: Using an interest-based recommendation model, the historical trajectories of the target user and other users are divided into sessions according to a specified time interval. Each session includes multiple points of interest. Based on the session, a global interaction hypergraph is generated, which includes nodes and hyperedges. Nodes represent points of interest, and hyperedges represent connections between points of interest in the same session. Generate global point of interest information based on the global interactive hypergraph; Based on the target user's initial information and global point of interest (POI) information, a prediction result of recommended target POIs to the target user is generated, including: generating spatiotemporal augmentation information of POIs based on global POI information and POI temporal information, wherein the POI temporal information includes the order in which the target user accesses POIs; generating spatiotemporal augmentation information of the target user based on the initial information of POIs in the target user's historical trajectory; and generating a prediction result of recommended target POIs to the target user based on the spatiotemporal augmentation information of POIs and the target user's spatiotemporal augmentation information. The interest point recommendation model is trained based on the predicted results of the target interest points recommended to the target user and the actual values ​​of the target interest points.

12. An interest point recommendation device, comprising: The trajectory processing module is configured to divide the historical trajectories of the target user and other users into sessions at specified time intervals, wherein a session includes multiple points of interest; The hypergraph generation module is configured to generate a global interactive hypergraph based on the session. The global interactive hypergraph includes nodes and hyperedges. Nodes represent points of interest, and hyperedges represent connections between points of interest in the same session. The global point of interest (POI) information generation module is configured to generate global point of interest information based on the global interactive hypergraph. The target point of interest (POI) generation module is configured to generate recommended POIs to the target user based on the target user's initial information and global POI information. This includes: generating spatiotemporal augmentation information for the POIs based on the global POI information and the POIs' temporal information, where the POIs' temporal information includes the order in which the target user accessed the POIs; generating spatiotemporal augmentation information for the target user based on the initial information of the POIs in the target user's historical trajectory; and generating recommended POIs to the target user based on the spatiotemporal augmentation information of the POIs and the target user's spatiotemporal augmentation information.

13. A training device for an interest point recommendation model, comprising: The trajectory processing module is configured to use an interest point recommendation model to divide the historical trajectories of the target user and other users into sessions according to a specified time interval. Each session includes multiple interest points. The hypergraph generation module is configured to generate a global interactive hypergraph based on the session. The global interactive hypergraph includes nodes and hyperedges. Nodes represent points of interest, and hyperedges represent connections between points of interest in the same session. The global point of interest (POI) information generation module is configured to generate global point of interest information based on the global interactive hypergraph. The target interest point generation module is configured to generate a prediction result of target interest points recommended to the target user based on the target user's initial information and global interest point information. This includes: generating spatiotemporal augmentation information of the interest points based on the global interest point information and the time information of the interest points, wherein the time information of the interest points includes information on the order in which the target user accesses the interest points; generating spatiotemporal augmentation information of the target user based on the initial information of the interest points in the target user's historical trajectory; and generating a prediction result of target interest points recommended to the target user based on the spatiotemporal augmentation information of the interest points and the target user's spatiotemporal augmentation information. The training module is configured to train an interest point recommendation model based on the predicted results of the target interest points recommended to the target user and the actual values ​​of the target interest points.

14. A spatiotemporal modeling system, comprising the interest point recommendation device according to claim 12 or the interest point recommendation model training device according to claim 13.

15. An electronic device comprising: Memory; as well as A processor coupled to the memory, the processor being configured to execute, based on instructions stored in the memory, the interest point recommendation method according to any one of claims 1 to 10 or the training method of the interest point recommendation model according to claim 11.

16. A computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the interest point recommendation method according to any one of claims 1 to 10 or the training method of the interest point recommendation model according to claim 11.

Images