A method for recommending points of interest based on users' spatio-temporal behavior and social information

Abstract

This invention discloses a method for recommending points of interest (POIs) based on user spatiotemporal behavior and social information. The method includes constructing a spatiotemporal knowledge graph based on user POI visit sequences and social relationships; learning POI representations in the spatiotemporal knowledge graph and constructing a POI transfer matrix; updating POI features through graph convolution using the POI transfer matrix as weights; calculating a user preference representation based on the user POI visit sequence and POI features; calculating the weights of historically visited POIs based on the spatiotemporal distances between POIs and the distances between POIs and user preferences; inputting the POI sequence into a recurrent neural network and updating a hidden representation using the weights of historically visited POIs; and connecting the hidden representation with the user preference representation, inputting the result into a recommendation model to generate the next POI that matches the user's preferences. This method can effectively model user spatiotemporal behavior, capturing behavioral patterns and user preferences, and making POI recommendations more accurate.

Description

Technical Field

[0001] The present invention relates to the technical field of data mining and recommendation systems, and in particular to a method for recommending points of interest based on user spatiotemporal behavior and social information. Background Art

[0002] In location-based social networks, users communicate with other users by sharing check-in information. The core content of check-in information is the check-in location, known as a point of interest (POI). With the rapid development of location-based social networks and the continuous expansion of social platforms, the number of POIs available to users has also increased dramatically. This often makes it difficult for users to quickly find POIs that match their preferences. To improve the efficiency of information screening and provide personalized services to users, recommender systems have emerged.

[0003] Existing technologies typically use continuous POI recommendation, predicting a user's next likely POI based on their historical POI visit sequence. Continuous POIs have strong temporal and spatial properties. This means that the user's likely choice in the current state is predicted based on their past check-in history. Furthermore, because check-ins require the user to physically visit the location, varying distances also influence the user's choice. Therefore, continuous POI recommendation possesses temporal and spatial characteristics.

[0004] Continuous POI recommendation requires not only considering users' personal preferences but also the temporal and spatial order of their check-in sequences. However, due to the complex and varied patterns of user behavior sequences, it is difficult to accurately model them. Existing methods also struggle to identify correlations between user behavior sequences and the behaviors of their friends. This makes it difficult to effectively utilize social information in continuous POI recommendation.

[0005] Patent application publication number CN 114417166 A discloses a method for recommending continuous points of interest (POIs) based on behavior sequences and dynamic social influence. The method involves obtaining check-in behavior data for a target user and their friends to obtain a check-in behavior sequence; dividing the check-in behavior sequence into long-term and short-term categories, obtaining long-term preference and short-term intent vectors, and aggregating these to obtain behavior sequence influence results; and obtaining the correlation between the check-in behavior sequences of the target user and their friends and determining a contextual influence vector to obtain the dynamic social influence results between the user and their friends, thereby obtaining continuous POI recommendation results. This patent application comprehensively considers the impact of sequence information and social factors on a user's continuous visit behavior within a spatiotemporal context.

[0006] Currently, in the scenario of POI recommendation, it is usually necessary to obtain the POI visit sequence through the user's active check-in behavior. However, most users have few check-in records, which leads to a serious data sparsity problem when modeling.

[0007] Patent application publication number CN 115422441 A discloses a method for continuous point-of-interest (POI) recommendation based on social spatiotemporal information and user preferences. The method involves obtaining long-term and short-term preference features from a user's check-in sequence; obtaining geographic spatial and user social information features at each location in the check-in sequence; and summarizing and fusing the long-term and short-term preference features, geographic spatial features, and user social information features into a multi-semantic feature. This feature is then fed into a recommendation model to generate a list of POI recommendations tailored to the user's preferences. This invention addresses the sparsity of check-in data and enhances the user's social information.

[0008] While existing research on POI recommendation methods has established some foundations, overall, research methods for POI recommendation are still relatively limited and require further study. Furthermore, mining this implicit semantic information, which implicitly reflects user POI preferences and the transition relationships between POIs, is another major challenge in the POI recommendation field. Summary of the Invention

[0009] In response to the above-mentioned deficiencies in the existing technology, the present invention provides a method for recommending points of interest based on user spatiotemporal behavior and social information. This method can effectively model user spatiotemporal behavior, capture behavioral patterns and user preferences, and make the recommendation of points of interest more accurate.

[0010] A method for recommending points of interest based on user spatiotemporal behavior and social information, comprising the following steps:

[0011] (S100) constructing a spatiotemporal knowledge graph based on the user's interest point access sequence and social relationships;

[0012] (S200) learning the representation of each entity and relationship in the spatiotemporal knowledge graph using a knowledge graph embedding algorithm, and calculating the semantic similarity between the interest point representations according to a semantic similarity function;

[0013] (S300) constructing an interest point transfer matrix based on the semantic similarity between interest point representations;

[0014] (S400) encoding and embedding the user and the point of interest to obtain a point of interest feature vector, using the point of interest transfer matrix as a weight, and updating the point of interest feature vector using a graph convolution algorithm;

[0015] (S500) aggregating the feature vectors of the points of interest according to the user's point of interest visit sequence to obtain a user preference representation;

[0016] (S600) Calculating the weight of historically visited points of interest based on the spatiotemporal distance between the points of interest and the distance between the points of interest and the user's preferences;

[0017] (S700) Inputting the feature vector of the point of interest into the recurrent neural network according to the user's point of interest visit sequence to obtain the corresponding hidden state sequence, and then aggregating the historical hidden states into the current hidden state according to the weights of the historically visited points of interest;

[0018] (S800) The current hidden state and the user's interest preference representation are connected and input into the recommendation model to calculate the next point of interest recommended for the user.

[0019] Preferably, in step (S100), the step of constructing a spatiotemporal knowledge graph based on the user's interest point access sequence and social relationships is:

[0020] (S110) Each user and point of interest is represented as an entity, and two users are represented as friends as a relationship r f , user visits to interest points are represented as relations r v , after a user visits a point of interest and then visits another point of interest, it is represented as a relationship r t , the straight-line distance between two points of interest is less than the set threshold ΔL, which is expressed as the relationship r s ;

[0021] (S120) Each entity is represented as a vertex in the knowledge graph, and each relationship is represented as a directed edge in the knowledge graph, so as to construct a user-interest point spatiotemporal knowledge graph.

[0022] Preferably, in step (S200), the specific steps of using the knowledge graph embedding algorithm to learn the representation of each entity and relationship in the spatiotemporal knowledge graph and calculating the semantic similarity between the interest point representations according to the semantic similarity function are:

[0023] (S210) Use the Trans family of algorithms to learn the representation of each entity and relationship in the spatiotemporal knowledge graph;

[0024] (S220) Using semantic similarity function Calculate the semantic similarity between interest points, where l1 and l2 are the representations of interest points, r t For the relationship r t The representation of r s For the relationship r s , where e is the natural logarithm.

[0025] More preferably, in step (S210), the Trans series algorithm is TransE, TransH, TransR or TransD.

[0026] Preferably, in step (S300), the step of constructing the interest point transfer matrix according to the semantic similarity between the interest point representations is:

[0027] (S310) For each point of interest li , select all other points of interest that are related to l i Points of interest closest to the top k Will and l i The semantic similarity of is taken as the element of the interest point transfer matrix, that is, in is the interest point transfer matrix;

[0028] (S320) Select the largest element in each row of M to construct a diagonal matrix D, and use D to normalize M to obtain a new interest point transfer matrix T = D -1 M.

[0029] Preferably, in step (S400), the steps of encoding and embedding the user and the point of interest to obtain the point of interest feature vector, using the point of interest transfer matrix as the weight, and updating the point of interest feature vector using the graph convolution algorithm are as follows:

[0030] (S410) Use the one-hot algorithm to encode the interest point and the user to obtain the interest point one-hot vector u o and user one-hot vector l o , and then use the embedding matrix E to embed the one-hot vector input to obtain the interest point feature vector l = El o And user feature vector u=Eu o ;

[0031] (S420) Perform self-connection on the interest point transfer matrix to obtain T c =T+I, and then use the out-degree matrix to normalize to get the new interest point transfer matrix Where I is the identity matrix, D c T c The out-degree matrix of

[0032] (S430) Using the interest point transfer matrix as a weight, the interest point feature vector is updated using a graph convolution algorithm to obtain a new interest point feature vector, where: X={l1,l2,…,l n}and Represent the interest point feature vectors before and after updating respectively.

[0033] Preferably, in step (S500), the specific steps of aggregating the interest point feature vectors according to the user's interest point visit sequence to obtain the user interest preference representation are:

[0034] For each user, at each time step i, the feature vectors of the points of interest are aggregated according to their point of interest visit sequence to obtain the user interest preference representation where ΔT i,j Visit points of interest for history j and currently visited POI i The time interval between them is in days, and γ is the decay weight.

[0035] Preferably, in step (S600), the step of calculating the weight of historically visited points of interest based on the spatiotemporal distance between points of interest and the distance between the points of interest and user preferences is as follows:

[0036] (S610) Calculate the spatiotemporal distance between points of interest: Use the spatiotemporal similarity function to calculate the spatiotemporal similarity between the historically visited points of interest and the currently visited points of interest. For each time step i, use the spatiotemporal similarity function to calculate the spatiotemporal similarity between the historically visited points of interest l and the currently visited points of interest l. j and currently visited POI i The spatiotemporal similarity between them, the spatiotemporal similarity function is expressed as

[0037]

[0038] where ΔD i,j is the spatial distance between interest points, α and β are attenuation weights;

[0039] (S620) Calculate the distance between the interest point and the user preference: use the preference similarity function Calculate user preference representation p and historically visited points of interest Similarity of preferences between

[0040] (S630) Fusion of spatiotemporal similarity and preference similarity to obtain historically visited POI weights

[0041] Preferably, in step (S700), the steps of inputting the feature vector of the points of interest into the recurrent neural network according to the user's point of interest visit sequence to obtain the corresponding hidden state sequence, and then aggregating the historical hidden states into the current hidden state according to the weights of the historically visited points of interest are as follows:

[0042] (S710) The point of interest feature vector is converted into the point of interest feature vector according to the user's point of interest visit sequence. Input the recurrent neural network and get the corresponding hidden state sequence {h1,h2,…,h n};

[0043] (S720) Aggregate the historical hidden states into the current hidden state h according to the weights of the historically visited points of interest i , get the updated current hidden state

[0044] Preferably, in step (S800), the step of connecting the current hidden state and the user preference representation and inputting them into the recommendation model to calculate the next point of interest recommended for the user is:

[0045] Concatenate the current hidden state and the user preference representation and input them into the fully connected layer to obtain the predicted probability of each point of interest. Where W is the weight of the fully connected layer, Concate(·) represents the concatenation function, and among all POIs, the POI with the highest predicted probability is the next recommended POI.

[0046] Optionally, before step (S100), the following steps are further included:

[0047] For each user Constructing a user's point of interest visit sequence Each access record r = (u, t, l) indicates that user u visited point of interest l at time t, and the access records are sorted in ascending order according to timestamp t.

[0048] Optionally, after step (S800), the following steps are further included:

[0049] The model parameters are updated using the mini-batch gradient descent algorithm, and the loss function used is expressed as

[0050]

[0051] Where m is the number of samples in a small batch, and the next point of interest selected by user u at time step i is l k , and Points of interest l k and the predicted probabilities of other points of interest, σ is the activation function.

[0052] Compared with the prior art, the present invention has at least the following beneficial effects:

[0053] 1. A user-POI knowledge graph is constructed based on user POI visit sequences and social relationships. The relationships between users, users and POIs, and POIs are comprehensively modeled. The POI transfer matrix is constructed using the learned POI and relationship representations, effectively capturing the temporal and spatial relationships between POIs. This helps the model learn different features of POIs, such as their popular time, impact range, and transfer relationships, thus solving the problem of lack of POI information.

[0054] 2. Graph convolution is performed on the interest point transfer matrix to update the interest point feature vector, so that the interest point can capture the interest point transfer relationship and local topological features. A time decay coefficient is added when constructing user preferences, so that user preferences can more accurately reflect the current state.

[0055] 3. When using a recurrent neural network to aggregate historical hidden states into the current hidden state, the spatial and temporal distances between historically visited POIs and currently visited POIs are considered. This allows for the periodic repetition of user behavior when visiting POIs, the distance between POIs and user preferences, and the integration of user preferences to make recommendation results more personalized.

[0056] 4. The present invention solves the problem that it is difficult to effectively model user spatiotemporal behavior and difficult to capture behavioral patterns. It realizes multi-level feature extraction of user interest point visit sequences, increases the utilization of user and interest point semantic information, enhances the learning of periodic behavior patterns and geographic spatial features, accurately recommends the next interest point, and saves users time in retrieving information. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 This is a flow chart of the method for recommending points of interest based on user spatiotemporal behavior and social information in this embodiment;

[0058] Figure 2 is the user-interest point spatiotemporal knowledge graph in this embodiment;

[0059] Figure 3 Schematic diagram of the structure of the point of interest recommendation system based on user spatiotemporal behavior and social information in this embodiment. DETAILED DESCRIPTION

[0060] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0061] Before constructing the spatiotemporal knowledge graph in this embodiment, each user can be pre- Constructing a user's point of interest visit sequence Each access record r = (u, t, l) indicates that user u visited point of interest l at time t, and the access records are sorted in ascending order according to timestamp t.

[0062] like Figure 1 As shown in FIG, the method for recommending points of interest based on user spatiotemporal behavior and social information includes the following steps:

[0063] (S100) Constructing a spatiotemporal knowledge graph based on the user's interest point access sequence and social relationships.

[0064] (S110) Each user and points of interest Represented as an entity, users u1 and u2 are friends represented as a relationship (u1, r f ,u2), user u visits point of interest l and is represented by the relationship (u,r v ,l), the user visits the interest point l1 and then visits the interest point l2, which is represented by the relationship (l1,r t ,l2), the straight-line distance between interest point l1 and interest point l2 is less than ΔL, which is expressed as the relationship (l1,r s ,l2);

[0065] (S120) Each entity is represented as a vertex in the knowledge graph, and each relationship is represented as a directed edge in the knowledge graph, so as to construct a user-interest point spatiotemporal knowledge graph.

[0066] Figure 2 The user-interest spatiotemporal knowledge graph is shown, where users u1 and u2 are friends, and relationship r f For directed edges from u1 to u2 and from u2 to u1, the relationship r v is a directed edge from u to l, relation r t is a directed edge from l1 to l2, relationship r s There are directed edges from l2 to l3 and from l3 to l2.

[0067] (S200) A knowledge graph embedding algorithm is used to learn the representation of each entity and relationship in the spatiotemporal knowledge graph, and the semantic similarity between the interest point representations is calculated according to a semantic similarity function.

[0068] (S210) using the TransR algorithm to learn the representation of each entity and relationship in the spatiotemporal knowledge graph;

[0069] (S220) Using semantic similarity function Calculate the semantic similarity between interest points, where l1 and l2 are the representations of interest points, r t For the relationship r t The representation of r s For the relationship r s , where e is the natural logarithm.

[0070] (S300) Constructing an interest point transfer matrix according to the semantic similarity between interest point representations.

[0071] (S310) For each point of interest l i , select all other points of interest that match l i Points of interest closest to the top k Will and li The semantic similarity of is taken as the element of the interest point transfer matrix, that is, in is the interest point transfer matrix;

[0072] (S320) Select the largest element in each row of M to construct a diagonal matrix D, and use D to normalize M to obtain a new interest point transfer matrix T = D -1 M.

[0073] (S400) Encoding and embedding the user and the point of interest to obtain a feature vector of the point of interest, and using the point of interest transfer matrix as a weight, and using a graph convolution algorithm to update the feature vector of the point of interest.

[0074] (S410) Use the one-hot algorithm to encode the interest point and the user to obtain the interest point one-hot vector u o and user one-hot vector l o , and then use the embedding matrix E to embed the one-hot vector input to obtain the interest point feature vector l = El o And user feature vector u=Eu o ;

[0075] (S420) Perform self-connection on the interest point transfer matrix to obtain T c =T+I, and then use the out-degree matrix to normalize to get the new interest point transfer matrix Where I is the identity matrix, D c T c The out-degree matrix of

[0076] (S430) Using the interest point transfer matrix as a weight, the interest point feature vector is updated using a graph convolution algorithm to obtain a new interest point feature vector, where: X={l1,l2,…,l n}and Represent the interest point feature vectors before and after updating respectively.

[0077] (S500) Aggregating the interest point feature vectors according to the user's interest point visit sequence to obtain the user preference representation.

[0078] For each user, at each time step i, the feature vectors of the points of interest are aggregated according to their point of interest visit sequence to obtain the user interest preference representation where ΔT i,j Visit points of interest for history j and currently visited POI i The time interval between them is in days, and γ is the decay weight.

[0079] (S600) Using the spatiotemporal similarity function to calculate the spatiotemporal similarity between the historically visited points of interest and the currently visited points of interest, and then using the preference similarity function to calculate the preference similarity between the user preference representation and the historically visited points of interest, the spatiotemporal similarity and the preference similarity are integrated to obtain the weights of the historically visited points of interest.

[0080] (S610) For each time step i, use the spatiotemporal similarity function to calculate the historical visited interest points l j and currently visited POI i The spatiotemporal similarity between them, the spatiotemporal similarity function is expressed as

[0081]

[0082] where ΔD i,j is the spatial distance between interest points, α and β are attenuation weights;

[0083] (S620) Calculate the distance between the interest point and the user preference: use the preference similarity function Calculate user preference representation p and historically visited points of interest Similarity of preferences between

[0084] (S630) Fusion of spatiotemporal similarity and preference similarity to obtain historically visited POI weights

[0085] (S700) Inputting the interest point feature vector into the recurrent neural network according to the user's interest point visit sequence to obtain the corresponding hidden state sequence, and then aggregating the historical hidden states into the current hidden state according to the historical interest point visit weights.

[0086] (S710) The point of interest feature vector is converted into the point of interest feature vector according to the user's point of interest visit sequence. Input the gated recurrent unit network GRU and get the corresponding hidden state sequence {h1,h2,…,h n};

[0087] (S720) Aggregate the historical hidden states into the current hidden state h according to the weights of the historically visited points of interest i , get the updated current hidden state

[0088] (S800) The current hidden state and the user's interest preference representation are connected and input into the recommendation model to calculate the next point of interest recommended for the user.

[0089] Concatenate the current hidden state and the user preference representation and input them into the fully connected layer to obtain the predicted probability of each point of interest. Where W is the weight of the fully connected layer, and Concate(·) represents the concatenation function. Among all POIs, the POI with the highest predicted probability is the next recommended POI.

[0090] After the above steps are completed, the mini-batch gradient descent algorithm can still be used to update the model parameters. The loss function used is expressed as

[0091]

[0092] Where m is the number of samples in a small batch, and the next point of interest selected by user u at time step i is l k , and Points of interest l k and the predicted probabilities of other points of interest, σ is the activation function.

[0093] Figure 3 The diagram shows the structure of the POI recommendation system based on user spatiotemporal behavior and social information. The embedding layer uses the one-hot algorithm to encode the POI and user to obtain the one-hot vector l o and u o , and then use the embedding matrix E to embed the one-hot vector to obtain the interest point feature vector l = El o And user feature vector u=Eu o The interest point transfer matrix calculated by the graph convolution layer (GCN Layer) based on the spatiotemporal knowledge graph (Knowledge Graph) The interest point feature vector X is updated using the graph convolution algorithm to obtain a new interest point feature vector The RNN layer receives the new interest point feature vector Input, use the gated recurrent unit network GRU to get the hidden state sequence {h1,h2,…,h n On the other hand, the feature vectors of the points of interest are aggregated according to the user's point of interest visit sequence to obtain the user preference representation p. The aggregation layer calculates the weight ω of the historically visited points of interest based on the spatiotemporal similarity function g(·) and the preference similarity function h(·). Based on this weight, the historical hidden state is aggregated into the current hidden state to obtain the updated current hidden state. The MLP layer (MLPLayer) concatenates the current hidden state and the user preference representation, inputs it into the fully connected layer, and obtains the predicted probability of each point of interest.

[0094] It will be easily understood by those skilled in the art that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for recommending points of interest based on user spatiotemporal behavior and social information, characterized in that: The following steps are involved: (S100) constructing a spatiotemporal knowledge graph based on the user's interest point access sequence and social relationships; (S200) learning the representation of each entity and relationship in the spatiotemporal knowledge graph using a knowledge graph embedding algorithm, and calculating the semantic similarity between the interest point representations according to a semantic similarity function; (S300) constructing an interest point transfer matrix based on the semantic similarity between interest point representations; The steps of constructing the interest point transfer matrix based on the semantic similarity between interest point representations are as follows: (S310) For each point of interest l i , select all other points of interest that match l i Points of interest closest to the top k Will and l i The semantic similarity of is taken as the element of the interest point transfer matrix, that is, in is the interest point transfer matrix; (S320) Select the largest element in each row of M to construct a diagonal matrix D, and use D to normalize M to obtain a new interest point transfer matrix T = D -1 M; (S400) encoding and embedding the user and the point of interest to obtain a point of interest feature vector, using the point of interest transfer matrix as a weight, and updating the point of interest feature vector using a graph convolution algorithm; (S500) aggregating the feature vectors of the points of interest according to the user's point of interest visit sequence to obtain a user preference representation; (S600) Calculating the weight of historically visited points of interest based on the spatiotemporal distance between the points of interest and the distance between the points of interest and the user's preferences; The steps of calculating the weights of historically visited points of interest based on the spatiotemporal distances between points of interest and the distances between points of interest and user preferences are as follows: (S610) Calculate the spatiotemporal distance between points of interest: Use the spatiotemporal similarity function to calculate the spatiotemporal similarity between the historically visited points of interest and the currently visited points of interest. For each time step i, use the spatiotemporal similarity function to calculate the spatiotemporal similarity between the historically visited points of interest l and the currently visited points of interest l. j and currently visited POI i The spatiotemporal similarity between them, the spatiotemporal similarity function is expressed as where ΔD i,j is the spatial distance between interest points, α and β are attenuation weights; (S620) Calculate the distance between the interest point and the user preference: use the preference similarity function Calculate user preference representation p and historically visited points of interest Similarity of preferences between (S630) Fusion of spatiotemporal similarity and preference similarity to obtain historically visited POI weights (S700) Inputting the feature vector of the point of interest into the recurrent neural network according to the user's point of interest visit sequence to obtain the corresponding hidden state sequence, and then aggregating the historical hidden states into the current hidden state according to the weights of the historically visited points of interest; (S800) The current hidden state and the user's interest preference representation are connected and input into the recommendation model to calculate the next point of interest recommended for the user.

2. The method according to claim 1, characterized in that In step (S100), the steps of constructing a spatiotemporal knowledge graph based on the user's interest point access sequence and social relationships are as follows: (S110) Each user and point of interest is represented as an entity, and two users are represented as friends as a relationship r f , user visits to interest points are represented as relations r v , after a user visits a point of interest and then visits another point of interest, it is represented as a relationship r t , the straight-line distance between two points of interest is less than the set threshold ΔL, which is expressed as the relationship r s ; (S120) Each entity is represented as a vertex in the knowledge graph, and each relationship is represented as a directed edge in the knowledge graph, so as to construct a user-interest point spatiotemporal knowledge graph.

3. The method according to claim 1, characterized in that In step (S200), the specific steps of using the knowledge graph embedding algorithm to learn the representation of each entity and relationship in the spatiotemporal knowledge graph and calculating the semantic similarity between the interest point representations according to the semantic similarity function are as follows: (S210) Use the Trans family of algorithms to learn the representation of each entity and relationship in the spatiotemporal knowledge graph; (S220) Using semantic similarity function Calculate the semantic similarity between interest points, where l1 and l2 are the representations of interest points, r t For the relationship r t The representation of r s For the relationship r s , where e is the natural logarithm.

4. The method according to claim 1, wherein In step (S400), the steps of encoding and embedding the user and the point of interest to obtain the point of interest feature vector, using the point of interest transfer matrix as the weight, and using the graph convolution algorithm to update the point of interest feature vector are as follows: (S410) Use the one-hot algorithm to encode the interest point and the user to obtain the interest point one-hot vector u o and user one-hot vector l o , and then use the embedding matrix E to embed the one-hot vector input to obtain the interest point feature vector l = El o And user feature vector u=Eu o ; (S420) Perform self-connection on the interest point transfer matrix to obtain T c =T+I, and then use the out-degree matrix to normalize to get the new interest point transfer matrix Where I is the identity matrix, D c T c The out-degree matrix of (S430) Using the interest point transfer matrix as a weight, the interest point feature vector is updated using a graph convolution algorithm to obtain a new interest point feature vector, where: X={l1,l2,…,l n }and Represent the interest point feature vectors before and after updating respectively.

5. The method according to claim 1, wherein In step (S500), the specific steps of aggregating the interest point feature vectors according to the user's interest point visit sequence to obtain the user's interest preference representation are: For each user, at each time step i, the feature vectors of the points of interest are aggregated according to their point of interest visit sequence to obtain the user interest preference representation where ΔT i,j Visit points of interest for history j and currently visited POI i The time interval between them is in days, and γ is the decay weight.

6. The method according to claim 1, characterized in that In step (S700), the steps of inputting the interest point feature vector into the recurrent neural network according to the user's interest point visit sequence to obtain the corresponding hidden state sequence, and then aggregating the historical hidden states into the current hidden state according to the historical interest point visit weights are as follows: (S710) The point of interest feature vector is converted into the point of interest feature vector according to the user's point of interest visit sequence. Input the recurrent neural network and get the corresponding hidden state sequence {h1,h2,…,h n }; (S720) Aggregate the historical hidden states into the current hidden state h according to the weights of the historically visited points of interest i , get the updated current hidden state 7. The method according to any one of claims 1 to 6, characterized in that Before step (S100), the process also includes: Constructing a user's point of interest visit sequence Each access record r = (u, t, l) indicates that user u visited point of interest l at time t, and the access records are sorted in ascending order according to timestamp t.

8. The method according to any one of claims 1 to 6, characterized in that After step (S800), the model parameters are updated using a mini-batch gradient descent algorithm, and the loss function used is Where m is the number of samples in a small batch, and the next point of interest selected by user u at time step i is l k , and Points of interest l k and the predicted probabilities of other points of interest, σ is the activation function.

Images