A Method for Point-of-Interest Recommendation Based on a Dual-Channel Neural Network with Spatiotemporal Attention

Abstract

The present invention discloses a method for recommending points of interest based on a dual-channel neural network with spatio-temporal attention. The architecture of the dual-channel neural network is adopted. The user group is used as the input to extract the time preference features and location preference features of the user group, and the individual user is used as the input to extract the personalized preference features of the user. The trajectory intention of the user is learned through a graph neural network, and personalized recommendations are made for the user by combining the time and location preference features. This method can learn the correlation between non-adjacent positions and non-consecutive visits. A suitable next point of interest is matched for the user through the user time preference and point of interest time preference matching layer, and it is more effective in the face of the cold start problem. When calculating the loss in the sampling method of the present invention, the sampling loss still has a momentum effect on the correct prediction, and all negative samplings have the opportunity to participate in the calculation of the cross-entropy loss. The present invention is more effective on multiple data sets and performs outstandingly when the sequence is short.

Description

Technical Field

[0001] The present invention relates to the technical field of point - of - interest recommendation systems, and more particularly to a point - of - interest recommendation method based on a dual - channel neural network with spatio - temporal attention. Background Art

[0002] The point - of - interest recommendation system is a branch of the recommendation system. Different from the traditional recommendation system methods, the point - of - interest recommendation system is related to the real geography and is closely affected by the geographical location. As Tobler's first law of geography states, "Everything is related to everything else, but near things are more related than distant things." In the point - of - interest recommendation system, Tobler's first law of geography means that users prefer to visit nearby locations rather than distant locations, and users may be interested in the points of interest around the points of interest they like. Geographical influence is the most important feature that differentiates the point - of - interest recommendation system from the traditional recommendation system and seriously affects users' access behaviors. In addition, in the traditional recommendation system, users usually express their preferences by explicitly providing ratings for items (such as books, movies, music, etc.), which are converted into a user - item rating matrix. The ratings are usually numerical and belong to a numerical range. The higher the rating, the higher the satisfaction. Different from the traditional recommendation system, users' preferences are reflected by the check - in frequency of locations, which is usually converted into a user - location check - in frequency matrix. Compared with ratings, the range of frequency data is larger. For example, a user may check in at some locations thousands of times, while at other locations, the user may check in only a few times. In addition, the sparsity of user - location checks in the frequency matrix is significantly higher than that of the user - item rating matrix, which brings greater challenges to point - of - interest recommendation. Additionally, the traditional recommendation system combines social relationships with ratings based on the assumption that friends are more likely to share more common interests and users are more likely to seek advice from friends to improve the recommendation quality. Some studies have shown that social relationships are beneficial to the recommendation system. However, in the point - of - interest recommendation system, previous studies have shown that approximately 96% of users share less than 10% of common visit interests, indicating that a large number of friends do not share anything in terms of POIs. Therefore, the influence of social influence on users' check - in behaviors is limited.

[0003] In recent years, more and more new methods have been proposed for the point - of - interest recommendation system. STAN is a state - of - the - art model that uses a two - layer attention mechanism architecture to aggregate gated spatio - temporal correlations in user trajectories. By using the relevant spatio - temporal information of all check - ins of the self - attention layer along the trajectory, this improvement allows for point - to - point interactions between non - adjacent locations and non - consecutive check - ins and has a clear spatio - temporal effect. However, there are still the following problems: 1) The preference characteristics of users at a specific time are not learned. 2) The existing hierarchical grids used for spatial discretization are insensitive to spatial distances. 3) Most of the previous work has not considered personalized recommendations for users. 4) The spatial connections between points of interest are often ignored. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to increase the recommendation accuracy of the model by learning the correlation between non - adjacent positions and non - continuous accesses. The present invention proposes a method for recommending points of interest based on a two - channel neural network with spatio - temporal attention.

[0005] To achieve the above object, the present invention provides a method for recommending points of interest based on a two - channel neural network with spatio - temporal attention, and the specific steps are as follows:

[0006] Step S1: Data pre - processing to obtain a data set, such that each user has more than a set number of records as input. Each record includes the user name, the name of the point of interest, and the access time. The data storing the latitude and longitude information of all points of interest is used as another input to the two - channel neural network model;

[0007] Step S2: Perform an embedding representation on the points of interest;

[0008] Step S3: Input the data set in Step S1 into the time preference mining and matching module. By extracting the access frequency features of each user at different time stamps and the frequency features of each location being accessed at different time stamps, obtain the feature of the preference degree of each user for each point of interest. In the time preference mining and matching module, by extracting the access frequency features of each user at different time stamps and the frequency features of each location being accessed at different time stamps, obtain the feature of the preference degree of each user for each point of interest;

[0009] Step S4: Input the records of a single user into the trajectory intention mining module. Through a graph neural network, establish a directed graph for the user's trajectory, propagate the information between points of interest through the directed graph, continuously update the features of the points of interest, and extract the feature of the user's trajectory intention through a layer of attention mechanism;

[0010] Step S5: Add the embedding representation of the preference degree feature of each user for each point of interest obtained in Step S3 to the embedding representation of the user's trajectory intention feature obtained in Step S4 to obtain the embedding representation of the user's final preference feature, and multiply it by the embedding representation of the point of interest obtained in Step S2 to obtain the score preference of the user for each point of interest;

[0011] Step S6: Input the single user in Step S4 into the multi - modal embedding module. At the same time, input the latitude and longitude of the point of interest into this module. Through the attention aggregation and matching mechanism, obtain the score of the user for each point of interest;

[0012] Step S7: Add the scores in Step S5 and Step S6 to obtain the final score, and calculate the loss through a balanced sampling;

[0013] Step S8: Repeat steps S3 to S7 until the loss obtained in step S7 no longer decreases, the network parameters are updated, and the model ends training;

[0014] Step S9: Input the user's access sequence into the trained model finally obtained in step S8, perform recommendation prediction, output the finally obtained recommendation result, and complete the recommendation.

[0015] Preferably, step S1 is specifically as follows: Select a suitable LBSN public dataset and group it to meet the requirements of the model for the data format; First, extract the user name, the name of the visited point of interest, and the time of visiting the point of interest from each record in the dataset, and merge them by user name to obtain a triple data composed of the user name, the sequence of the user's visited points of interest, and the sequence of the user's visit times; During the training process, except for the first point of interest, each point of interest will become a label once. At this time, all the points of interest in front of it are used as variables, and the points of interest behind it are not visible during training to ensure the time sequence. The last visited point of interest of each user is used as the test label, and all the points of interest in front of it are used as variables.

[0016] Preferably, the time features of both the user and the point of interest are extracted simultaneously through the user's overall behavior records, and the time preferences of the two are matched to obtain the user's preference degree for the point of interest.

[0017] Preferably, in the time preference mining module, the time stamp sequence of the user's check-in is taken modulo 24, the number of hours in a day, to obtain a 24-dimensional check-in frequency feature of user u i ;

[0018] Then, the time stamp is divided by 24 and taken modulo 7, the number of days in a week, to obtain a 7-dimensional check-in frequency feature of user u i ;

[0019] The 24-dimensional check-in frequency feature and the 7-dimensional check-in frequency feature are concatenated to obtain a 31-dimensional time preference feature u i of user u i t ;

[0020] The check-in time preference feature I j of the point of interest I j t is obtained by the same method;

[0021] The set of time preference features of all users is denoted as U t , and the set of time preference features of all points of interest is denoted as L t ;

[0022] In the time preference matching module, the user is matched with the point of interest through the following algorithm:

[0023] U t ′=tanh(U t )

[0024] L t ′=tanh(L t )

[0025]

[0026] preference(Q, K, V)=sum(softmax(QK T )V)

[0027] Among them, tanh is the activation function, E(l) is the embedded representation of the point of interest. Supplement the meaning of the parameters when they first appear. For example, d is the dimension of E(l), |U| is the total number of users in the user set, and Q, K, and V represent the query, key, and value respectively.

[0028] Preferably, step S4 is specifically as follows:

[0029] Input the access sequence of a single user into the trajectory intention discovery module, and update the set E(l) of the embedded representations of the points of interest through the constructed trajectory graph A. When there is a continuous access record between two points of interest, it is considered that there is a path between the two points of interest, and the out-degree graph and in-degree graph of the two points of interest are obtained. After splicing the two, the trajectory graph A for calculation is obtained, and the embedded representations of the nodes in the graph are continuously updated through a gated graph neural network. The specific algorithm is as follows:

[0030]

[0031]

[0032]

[0033]

[0034]

[0035] Among them, v is the embedded representation of each node in the trajectory graph, that is, the element in E(l), and W, b, and U are the parameters to be learned. A v is the embedded representation of the trajectory graph A;

[0036] Calculate the embedded representation of the user's access intention through an attention mechanism. The specific algorithm is as follows:

[0037] α i =q T(Sigmoid(W p v′ i +b p ))

[0038]

[0039] Among them, q, W p , b p are learnable weights, α is the preference value of the user for each point of interest after learning, and p is the embedded representation of the user's access intention.

[0040] Preferably, step S6 is specifically as follows:

[0041] Obtain the score of the user for each point of interest through the attention aggregation and matching mechanism;

[0042] First, perform attention aggregation between the points of interest. The specific algorithm is as follows:

[0043] S(u) = Attention(E(u)W Q , E(u)W K , E(u)W V , E(Δ), M)

[0044]

[0045] Among them, E(u) is the embedded representation of the user, E(Δ) is the embedded representation of the spatial distance between the points of interest, and W Q , W K , W V are learnable parameters.

[0046] M is a mask matrix used to ensure that the sequences in the front in the time series cannot see the sequence information in the back during learning;

[0047] Then, perform matching of the attention mechanism on the aggregated points of interest. The specific algorithm is as follows:

[0048] A(u) = Matching(E(l), S(u), E(V))

[0049]

[0050] Among them, E(l) is the embedded representation of all points of interest, and E(N) is the embedded representation of the candidate points of interest.

[0051] Preferably, the balanced sampling loss calculation method described in step S7 is as follows:

[0052]

[0053] Where k represents the subscript of the label, and s represents the s samples selected. When s is equal to L, all the points of interest are selected as samples, and at this time, the balanced sampling loss is the cross-entropy loss.

[0054] Therefore, the beneficial effects of the present invention are as follows: The dual-channel model is used to extract the group spatio-temporal features and individual spatio-temporal features of users respectively, learn the behavior preferences of users at specific times, and combine the graph neural network to provide personalized recommendations for users, improving the recommendation accuracy of the model and enabling the model to converge faster. Moreover, the present invention performs more prominently with a shorter sequence length.

[0055] Next, through the drawings and embodiments, the technical solutions of the present invention will be further described in detail. Description of the Drawings

[0056] Figure 1 is the overall system structure of the present invention;

[0057] Figure 2 is the detailed structure diagram of the present invention;

[0058] Figure 3 is the structure diagram of the trajectory intention mining module. Specific Embodiments

[0059] Embodiment

[0060] Next, in conjunction with the drawings, the embodiments of the present invention will be described in detail.

[0061] A method for recommending points of interest based on a dual-channel neural network with spatio-temporal attention, the specific steps are as follows:

[0062] Step S1: Data preprocessing to obtain user access records.

[0063] Data preprocessing obtains a data set such that each user has more than a set number of records as input. Each record includes the user name, the name of the point of interest, and the access time. The data storing the longitude and latitude information of all points of interest is used as another input to the dual-channel neural network model.

[0064] Select an appropriate LBSN public dataset and group it to meet the requirements of the model for data format. First, extract the user name, the name of the visited point of interest, and the time of visiting the point of interest from each record in the dataset, and merge them by user name to obtain a triple data consisting of the user name, the sequence of the user's visited points of interest, and the sequence of the user's visit times. During the training process, except for the first point of interest, each point of interest will become a label once. At this time, all the points of interest before it are used as variables, and the points of interest after it are not visible during training to ensure the time sequence. The last visited point of interest of each user is used as the test label, and all the points of interest before it are used as variables.

[0065] As Figure 1 shown, the overall structure of the present invention is presented. The data preprocessing part is at the entrance of the structure of the present invention and is responsible for preliminarily processing the original data to form the data structure required by the prediction model.

[0066] Step S2: Perform an embedding representation on the points of interest. Use a d-dimensional vector E(l) to represent the embedding of the point of interest. In subsequent calculations, the embedding expression of the point of interest will be continuously updated through calculations.

[0067] Step S3: Input the dataset in Step S1 into the temporal preference mining and matching module. By extracting the access frequency features of each user at different timestamps and the access frequency features of each location at different timestamps, the feature of the preference degree of each user for each point of interest is obtained. In the temporal preference mining and matching module, by extracting the access frequency features of each user at different timestamps and the access frequency features of each location at different timestamps, the feature of the preference degree of each user for each point of interest is obtained.

[0068] First, obtain the temporal preference features of the user and the temporal preference features of the points of interest by traversing all user records, and obtain the preference degree of the user for each location through the temporal preference matching module.

[0069] As Figure 2 shown, the overall structure of the present invention is presented, where the Temporal Preference Discovering and Matching Module is divided into a temporal preference mining module and a temporal preference matching module.

[0070] In the temporal preference mining module, take the remainder of the timestamp sequence of the user's check-in with respect to the number of hours in a day, 24, to obtain a 24-dimensional check-in frequency feature of user u i Then divide the timestamp by 24 and take the remainder with respect to the number of days in a week, 7, to obtain user u iA 7-dimensional check-in frequency feature. Concatenate the 24-dimensional check-in frequency feature and the 7-dimensional check-in frequency feature to obtain the user u i A 31-dimensional time preference feature u of the user i t The same method is used to obtain the check-in time preference feature I of the point of interest I j j t Denote the set of time preference features of all users as U t Denote the set of time preference features of all points of interest as L t

[0071] In the time preference matching module, match users with points of interest through the following algorithm:

[0072] U t ′ = tanh(U t )

[0073] L t ′ = tanh(L t )

[0074]

[0075] preference(Q, K, V) = sum(softmax(QK T ))V

[0076] where tanh is the activation function, E(l) is the embedded representation of the point of interest. Supplement the meaning of the parameters for the first appearance, for example, d is the dimension of E(l), |U| is the total number of users in the user set, and Q, K, and V represent the query, key, and value respectively.

[0077] Step S4: Input the record of a single user into the trajectory intention mining module. Through a graph neural network, establish a directed graph for the user's trajectory, propagate the information between points of interest through the directed graph, continuously update the features of the points of interest, and extract the features of the user's trajectory intention through a layer of attention mechanism.

[0078] Input the access sequence of a single user into the trajectory intention discovery module. Update the set E(l) of the embedded representations of the points of interest through the constructed trajectory graph A. As Figure 3 shown, when there is a continuous access record between two points of interest, it is determined that there is a path between the two points of interest, and the out-degree graph and in-degree graph of the two points of interest are obtained. After splicing them, the trajectory graph A for calculation is obtained, and the embedded representations of the nodes in the graph are continuously updated through a gated graph neural network. The specific algorithm is as follows:

[0079] ​

[0080]

[0081]

[0082]

[0083]

[0084] Among them, v is the embedded representation of each node in the trajectory graph, that is, an element in E(l), and W, b, and U are parameters to be learned. A v is the embedded representation of the trajectory graph A.

[0085] Finally, an attention mechanism is used to calculate the embedded representation of the user's access intention. The specific algorithm is as follows:

[0086] α i = q T (Sigmoid(W p v′ i + b p ))

[0087]

[0088] Among them, q, W p , b p are learnable weights, α is the preference value of the user for each point of interest after learning, and p is the embedded representation of the user's access intention.

[0089] Step S5: Add the embedded representation of the preference degree feature of the user for each point of interest obtained in step S3 to the embedded representation of the user trajectory intention feature obtained in step S4 to obtain the embedded representation of the user's final preference feature. The specific algorithm is as follows:

[0090] p′ = p i v + p i t .

[0091] Multiply it by the embedded representation of the point of interest obtained in step S2 to obtain the score preference of the user for each point of interest.

[0092] Step S6: Input a single user in step S4 into the multi-modal embedding module. At the same time, input the longitude and latitude of the point of interest into this module. Through the attention aggregation and matching mechanism, obtain the score of the user for each point of interest.

[0093] Through the attention aggregation and matching mechanism, obtain the score of the user for each point of interest;

[0094] First, perform attention aggregation between points of interest. The specific algorithm is as follows:

[0095] S(u) = Attention(E(u)W Q , E(u)W K , E(u)W V , E(Δ), M)

[0096]

[0097] Among them, E(u) is the embedding representation of the user, E(Δ) is the embedding representation of the spatial distance between points of interest, W Q 、W K 、W V are learnable parameters, and M is a mask matrix used to ensure that the previous sequences in the time series cannot see the subsequent sequence information during learning.

[0098] Then, perform matching of the attention mechanism on the aggregated points of interest. The specific algorithm is as follows:

[0099] A(u) = Matching(E(l), S(u), E(N))

[0100]

[0101] Among them, E(l) is the embedding representation of all points of interest, and E(N) is the embedding representation of the candidate points of interest.

[0102] Step S7: Add the scores in Step S5 and Step S6 to obtain the final score, and calculate the loss through a balanced sampling.

[0103] The score calculation is as follows:

[0104]

[0105] Among them, A(u i ) represents the score of user i for each point of interest.

[0106] The calculation method of the balanced sampling loss described in Step S7 is as follows:

[0107]

[0108] Among them, k represents the subscript of the label, s represents the s samples selected. When s is equal to L, all points of interest are selected as samples, and at this time, the balanced sampling loss is the cross-entropy loss.

[0109] Step S8: Repeat Step S3 to Step S7 until the loss obtained in Step S7 no longer decreases, the network parameters are updated, and the model ends training;

[0110] Step S9: Input the user's access sequence into the trained model finally obtained in Step S8 for recommendation prediction, output the finally obtained recommendation result, and complete the recommendation.

[0111] Extract the temporal features of the user and the point of interest simultaneously based on the user's overall behavior record, and match their temporal preferences to obtain the user's preference degree for the point of interest.

[0112] The following table shows the experimental results of the method of the present invention and the existing methods on four datasets.

[0113]

[0114] The measurement criterion is recall@k, that is, in the case of selecting k candidates, if the label is among these k samples, this is regarded as a hit, and calculate the proportion of the users with hits among all users. Compared with other methods, our method has at least a 10% improvement.

[0115] The following table shows that the method of the present invention performs more prominently when the sequence length is shorter, which means that our method is more effective in the face of cold start.

[0116]

[0117] Therefore, the dual-channel model is used to extract the group spatio-temporal features and individual spatio-temporal features of the user respectively, learn the user's behavior preferences at a specific time, and combine the graph neural network to provide personalized recommendations for the user, improving the recommendation accuracy of the model and enabling the model to converge faster, and the present invention performs more prominently on a shorter sequence length.

[0118] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that they can still modify or equivalently replace the technical solutions of the present invention, and these modifications or equivalent replacements cannot make the modified technical solutions deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for recommending points of interest based on a dual-channel neural network with spatio-temporal attention, characterized in that, The specific steps are as follows: Step S1: Preprocess the data to obtain a dataset such that each user has more than a set number of records as input. Each record includes the user name, the name of the point of interest, and the access time. Use the data containing the longitude and latitude information of all points of interest as another input to the dual-channel neural network model; Step S2: Perform an embedding representation on the points of interest; Step S3: Input the dataset in Step S1 into the time preference mining and matching module. By extracting the access frequency features of each user at different timestamps and the access frequency features of each location at different timestamps, obtain the feature of the preference degree of each user for each point of interest. In the time preference mining and matching module, by extracting the access frequency features of each user at different timestamps and the access frequency features of each location at different timestamps, obtain the feature of the preference degree of each user for each point of interest; Step S4: Input the records of a single user into the trajectory intention mining module. Through a graph neural network, establish a directed graph for the user's trajectory. Propagate the information between points of interest through the directed graph, continuously update the features of the points of interest, and extract the feature of the user's trajectory intention through a layer of attention mechanism; Step S5: Add the embedding representation of the preference degree feature of each user for each point of interest obtained in Step S3 to the embedding representation of the user's trajectory intention feature obtained in Step S4 to obtain the embedding representation of the user's final preference feature. Multiply it by the embedding representation of the point of interest obtained in Step S2 to obtain the score preference of the user for each point of interest; Step S6: Input the single user in Step S4 into the multi-modal embedding module. At the same time, input the longitude and latitude of the point of interest into this module. Through the attention aggregation and matching mechanism, obtain the score of the user for each point of interest; Step S7: Add the scores in Steps S5 and S6 to obtain the final score, and calculate the loss through a balanced sampling; Step S8: Repeat Steps S3 to S7 until the loss obtained in Step S7 no longer decreases, the network parameters are updated, and the model ends training; Step S9: Input the user's access sequence into the trained model finally obtained in Step S8, perform recommendation prediction, output the finally obtained recommendation result, and complete the recommendation.

2. The method for recommending points of interest of a dual-channel neural network based on spatio-temporal attention according to claim 1, wherein Step S1 is specifically as follows: Select a suitable LBSN public dataset and group it to meet the requirements of the model for the data format; First, extract the user name, the name of the accessed point of interest, and the access time of the point of interest from each record in the dataset, and merge them according to the user name to obtain a triple data composed of the user name, the sequence of the user's accessed points of interest, and the sequence of the user's access times; During the training process, except for the first point of interest, each point of interest will become a label once. At this time, all the points of interest in front of it are used as variables, and the points of interest behind it are not visible during training to ensure the timeliness. The last accessed point of interest of each user is used as the test label, and all the points of interest in front of it are used as variables.

3. A method for recommending points of interest based on a dual-channel neural network with spatio-temporal attention according to claim 1, characterized in that, Extract the temporal features of both the user and the point of interest based on the user's overall behavior records, and match the temporal preferences of the two to obtain the user's preference degree for the point of interest.

4. The method for recommending a point of interest based on a dual-channel neural network with spatio-temporal attention according to claim 3, characterized in that In the time preference mining module, the time stamp sequence of the user's check-in is modulo 24 of the number of hours in a day, obtaining a 24-dimensional check-in frequency feature of user u i ; Then divide the timestamp by 24 and take the remainder when divided by 7, the number of days in a week, to obtain a 7-dimensional check-in frequency feature of user u i ; Concatenate the 24-dimensional check-in frequency feature and the 7-dimensional check-in frequency feature to obtain a 31-dimensional time preference feature u of user u i ; i t ; The same method is used to obtain the point of interest I j The check-in time preference feature I j t ; Denote the set of time preference features of all users as U t and denote the set of time preference features of all points of interest as L t ; In the temporal preference matching module, the user and the point of interest are matched through the following algorithm: U t′ =tanh(U t ) L t′ = tanh(L t ) preference(Q, K, V) = sum(softmax(QK T )V) Where tanh is the activation function, E(l) is the embedded representation of the point of interest, d is the dimension of E(l), |U| represents the total number of users in the user set, and Q, K, and V represent the query, key, and value respectively.

5. The method for recommending points of interest of a dual-channel neural network based on spatio-temporal attention according to claim 1, wherein Step S4 is specifically as follows: Input the access sequence of a single user into the trajectory intention discovery module, and update the set E(l) of the embedded representations of the points of interest through the constructed trajectory graph A. When there is a continuous access record between two points of interest, it is determined that there is a path between the two points of interest, and the out-degree graph and in-degree graph of the two points of interest are obtained. After splicing the two, the trajectory graph A for calculation is obtained, and the embedded representations of the nodes in the graph are continuously updated through a gated graph neural network. The specific algorithm is as follows: Among them, v is the embedding representation of each node in the trajectory graph, that is, an element in E(l), and W, b, and U are parameters to be learned, and A v is the embedding representation of the trajectory graph A; Calculate the embedded representation of the user's access intention through an attention mechanism. The specific algorithm is as follows: α i = q T (Sigmoid(W p v′ i + b p )) Among them, q, W p , b p are learnable weights, α is the preference value of the user for each point of interest after learning, and p is the embedded representation of the user's access intention.

6. The method for recommending points of interest of a dual-channel neural network based on spatio-temporal attention according to claim 1, wherein Step S6 is specifically as follows: Obtain the score of the user for each point of interest through the attention aggregation and matching mechanism; First, perform attention aggregation between the points of interest. The specific algorithm is as follows: S(u) = Attention(E(u)W Q , E(u)W K , E(u)W V , E(Δ), M) Among them, E(u) is the embedded representation of the user, E(Δ) is the embedded representation of the spatial distance between points of interest, and W Q , W K , W V are learnable parameters, and M is a mask matrix used to ensure that the previous sequences in the time series cannot see the information of the subsequent sequences during learning; Then, perform the matching of the attention mechanism on the aggregated points of interest. The specific algorithm is as follows: A(u) = Matching(E(l), S(u), E(N)) Where E(l) is the embedded representation of all points of interest, and E(N) is the embedded representation of the candidate points of interest.

7. A method for recommending points of interest based on a dual-channel neural network with spatio-temporal attention according to claim 1, characterized in that, The calculation method of the balanced sampling loss in step S7 is as follows: Where k represents the subscript of the label, s represents the s samples selected. When s is equal to L, all points of interest are selected as samples, and at this time, the balanced sampling loss is the cross-entropy loss.

Images