POI recommendation method fusing social relation network and spatio-temporal context information

Abstract

The invention belongs to the technical field of big data analysis, and particularly relates to a POI recommendation method fusing social relation network and spatio-temporal context information. In order to solve the problems that the influence of social relations on personal access tracks is not fully considered, and spatio-temporal context information features are not effectively extracted and fused, user features and time information features are subdivided into multiple sub-features, the relations between the features are deeply mined, meanwhile, a regional user track graph is constructed to model a POI-POI social relation network, and the POI-POI social relation network is established. User and spatio-temporal context information features are comprehensively learned through an encoder, weighted fusion is carried out, and dynamic preferences of users are captured, so that the accuracy of POI prediction is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of big data analysis, and in particular relates to a POI recommendation method integrating social relationship network and spatiotemporal context information. Background Art

[0002] With the popularity of mobile devices and the rapid development of Internet technology, mobile Internet has become an indispensable part of people's lives and work. This trend has led to an explosive growth in data volume, which has brought inconvenience to life. In this context, Point of Interest (POI) recommendation has become one of the hot topics in location-based social network research. It not only helps users quickly filter out places of interest from complex information, but also helps companies find potential customers, which has very important practical significance.

[0003] As a branch of the recommendation system, POI recommendation mainly relied on cumbersome manual feature extraction and traditional machine learning-based methods in the early days, which limited its efficiency and recommendation results. In recent years, it has become an obvious trend to introduce deep learning into the POI recommendation task. Deep learning methods can effectively model the complex relationship between users and POIs, automatically extract deep features from data, and capture the dynamic changes of user behaviors and interests, such as recurrent neural networks (RNN), transformers, and graph neural networks (GNN). Li et al. redefined social networks and incorporated features extracted from the check-in information of users' friends into the matrix decomposition model, which alleviated the cold start problem to a certain extent. Wang et al. proposed the CPAM recommendation algorithm, which uses LSTM with a multi-level attention mechanism to obtain user preferences, fully considers the different contextual information weights of each check-in, and the different effects of each check-in on the next POI visited. Yu et al. proposed a socially aware self-supervised triple training framework SEPT, which uses user social information to expand the user data view and improves recommendations by discovering self-supervisory signals from two complementary perspectives of the original data.

[0004] Existing research has achieved many results and progress, but there are still problems that need to be further studied and solved: Traditional POI recommendation methods do not consider the spatiotemporal characteristics of user check-in sequences and cannot effectively model the sequences. POI recommendation methods based on user information features will lead to a trade-off between global and local relationships when extracting features when the user's visit records are very long. POI recommendation methods based on spatiotemporal information still need to be improved in terms of data processing quality and comprehensiveness of feature extraction when mining the spatiotemporal information implicit in user check-in data. Summary of the invention

[0005] In order to solve the problems that the impact of social relations on individual visit trajectories is not fully considered and the spatiotemporal context information features are not effectively extracted and integrated, the present invention provides a POI recommendation method that integrates social relationship networks and spatiotemporal context information.

[0006] To achieve the above object, the present invention adopts the following technical solutions:

[0007] A POI recommendation method integrating social relationship network and spatiotemporal context information of the present invention comprises the following steps:

[0008] Step 1: Obtain user sign-in sequence data, perform data preprocessing, and construct a regional user trajectory map;

[0009] Step 2: Input the user check-in sequence data preprocessed in step 1 and the constructed domain user trajectory map into the pre-trained POI recommendation model to make the next POI recommendation; the POI recommendation model integrates the social relationship network and spatiotemporal context information, including a social relationship learning module, a POI spatiotemporal-category feature fusion module, an encoder and a decoder; the social relationship learning module performs feature aggregation and noise filtering on the regional user trajectory map through a graph convolutional neural network integrated with a gated recurrent unit, learns the features of each POI, and obtains the feature vector E of each POI. p The POI spatiotemporal-category feature fusion module performs preliminary feature extraction and fusion on the POI spatiotemporal information of different dimensions in the user check-in sequence data through the time-space fusion module and the time-category fusion module to obtain the time-space fusion feature E of the user check-in activity. l,ts and time-category fusion feature E c,t The encoder extracts features from the user trajectory sequence through the user feature learning module and the spatiotemporal feature learning module to obtain the user feature Z u and the spatiotemporal characteristics Z t ; The decoder is based on user feature Z u and the spatiotemporal characteristics Z t ,The next POI, the visiting time period and the POI category are predicted respectively through three multi-layer perceptrons.

[0010] Furthermore, the regional user trajectory graph in step 1 is defined as: a directed graph consisting of a node set V, an edge set E, a node attribute S and a weight W, wherein the node set V consists of the POIs that the user has visited historically, and the edge set E is a set of directed edges. When a user visits two POIs p1 and p2 consecutively within a period of time, a directed edge e is added between p1 and p2. The weight W of each directed edge is equal to the number of times the sub-trajectory appears in the regional trajectory graph, and the node attribute S includes the spatial coordinates, category and visit frequency of each POI.

[0011] Furthermore, the specific process of the social relationship learning module includes:

[0012] Step 2.1: Calculate the normalized Laplacian matrix of the regional user trajectory graph adjacency matrix

[0013] Step 2.2: Use a single-layer graph convolutional neural network to normalize the Laplacian matrix Perform feature extraction;

[0014] Step 2.3: Update the hidden state of each node using the gated recurrent unit;

[0015] Step 2.4: After repeating steps 2.2 and 2.3 multiple times, the obtained feature vector is compared with the feature vector x extracted by the first single-layer graph convolutional neural network. t Splice and merge to get the feature vector E of each POI p , the calculation process is expressed as follows:

[0016]

[0017] In the formula, H() represents a single-layer graph convolution operation, f T () represents a gated recurrent unit with T time steps, W G,C and b G,C denote weights and biases respectively, · denotes concatenation, is the Laplace matrix.

[0018] Furthermore, the specific process of the POI spatiotemporal-category feature fusion module includes:

[0019] 1) Time-space fusion module: Adopt the encoding strategy of the CTLE model to obtain the timestamp feature vector E ts ; Input the coordinates of each POI into the fully connected layer to obtain the spatial feature vector E l ; The timestamp feature vector E ts and the spatial eigenvector E l The splicing is performed and input into the fully connected layer for feature fusion to obtain the time-space fusion feature E of the user's check-in activity. l,ts , the calculation process is expressed as follows:

[0020] E l,ts =σ(W l,ts [E ts ·E l ]+b l,ts ) (2)

[0021] Where W l,ts represents the learnable weight vector, bl,ts represents the deviation, σ represents the nonlinear activation function LeakyReLU;

[0022] 2) Time-category fusion module: Adopt the encoding strategy of the time encoding model time2vector to encode the absolute access time and obtain the time feature vector E t ; Embed each POI category into a category feature vector E c Represented; the category feature vector E c and the time feature vector E t The splicing is then input into the fully connected layer for feature fusion to obtain the time-category fusion feature E of the user's check-in activity. c,t , the calculation process is expressed as follows:

[0023] E c,t =σ(W c,t [E c ·E t ]+b c,t ) (3)

[0024] Where W c,t represents the learnable weight vector, b c,t represents the bias, and σ represents the nonlinear activation function LeakyReLU.

[0025] Furthermore, the user trajectory sequence is defined as: consisting of the user's continuous check-in records in chronological order; after obtaining the time-space fusion feature representation E of the user's check-in activity l,ts And the time-category fusion feature representation E c,t After that, the embedding representation of user check-in q is divided into two parts: and Each user's trajectory sequence Both are embedded by two user check-in sequences and express.

[0026] Furthermore, the specific process of the encoder includes:

[0027] 1) User feature learning module: Construct the user feature vector E of the target user u’s personal information through an embedding layer u , construct the user check-in data into two modes: graph-based user trajectory graph and sequence-based user check-in sequence;

[0028] For the user trajectory graph, a graph convolutional neural network with a gated recurrent unit is used to extract features, and the obtained POI feature vector and user feature vector E uSplicing, input to the fully connected layer for nonlinear transformation, so as to obtain the graph-based POI-User fusion feature vector The calculation process is expressed as follows:

[0029]

[0030] in, and represent weight and bias respectively, represents splicing, σ L Represents the nonlinear activation function LeakyReLU;

[0031] For user check-in sequences, a lightweight self-attention mechanism and channel convolution are used to capture global context information and local context information respectively;

[0032] To capture global context information, firstly transform the feature vector E obtained by the social relationship learning module into p and user feature vector E u Splicing, output sequence-based POI-User fusion feature vector Then use a simplified lightweight self-attention mechanism to obtain global information, input it into the fully connected layer FC1 for nonlinear transformation, and obtain the sequence-based POI-User global feature vector The calculation process is expressed as follows:

[0033]

[0034] in, and Represent weights and biases respectively, Q, K, and V represent the input The same query weight matrix, key weight matrix and value weight matrix, d k represents the variance, T represents the transpose of the matrix, Represents the weight of the fully connected layer FC1, AvgPooling represents average pooling, and LayerNorm represents layer normalization;

[0035] To capture local context information, the sequence-based POI-User fusion feature vector The two-dimensional channel convolution layer conv2D is input to obtain local information, and then input to the fully connected layer FC2 for nonlinear transformation to obtain the sequence-based POI-User local feature vector The calculation process is expressed as follows:

[0036]

[0037] in, Represents the weight of the fully connected layer FC2;

[0038] After obtaining the graph-based POI-User fusion feature vector Sequence-based POI-User global feature vector and the local eigenvector After that, the user feature learning module outputs the user feature Z u , which is defined by the weighted sum of these three feature vectors, and two aggregation hyperparameters α1 and α2 are introduced to reasonably allocate the fusion ratio. The calculation process is expressed as follows:

[0039]

[0040] Among them, σ t Represents the activation function tanh;

[0041] 2) Spatiotemporal feature learning module: given a user trajectory sequence Embedding representation of two user check-in sequences and Will and As input tensors, they are respectively passed into the spatiotemporal feature learning module to capture spatiotemporal features of different dimensions. The module has a multi-layer structure. For the pth layer, the input and Through the transformation of multi-head self-attention layers, the calculation process is expressed as follows:

[0042]

[0043] Among them, W Q , W K , W V represent the query weight matrix, key weight matrix, and value weight matrix respectively, E ij represents the correlation between the i-th and j-th check-ins, α ij Indicates the influence weight of other check-in activities on the current check-in;

[0044] The output Head obtained by different attention heads i The layers are stacked and the representations from different attention spaces are merged through linear transformation. The encoder uses layer normalization LayerNorm and residual connection technology. The calculation process is expressed as follows:

[0045]

[0046] After the multi-head self-attention layer, the feedforward neural network FNN and activation function ReLU are used to give the model nonlinearity. The calculation process is expressed as follows:

[0047]

[0048] Among them, W1 and W2 represent weights, b1 and b2 represent biases, so the output of the encoder layer p is expressed as follows:

[0049]

[0050] After stacking the multi-head self-attention blocks multiple times, the output spatiotemporal feature Z of the spatiotemporal feature learning module is t Output features from time-space and time-category output features The weighted sum definition of , introduces two aggregation hyperparameters β1 and β2 to reasonably distribute the fusion ratio. The calculation process is expressed as follows:

[0051]

[0052] Among them, σ t represents the activation function tanh, p * Represents the last layer of the spatiotemporal feature learning module.

[0053] Furthermore, the decoder uses three multi-layer perceptrons to predict the next POI, the visit time period, and the POI category, respectively, as defined in the following formula:

[0054]

[0055] in, Indicates the next POI, Indicates the access time period. represents the POI category, W poi , W time , W cat Respectively represent their respective weights, b poi 、b time 、b cat Respectively represent their own deviations.

[0056] Going further, it also includes the training of the POI recommendation model, and its loss function is defined as follows:

[0057]

[0058] In the formula, represents the overall loss function of the POI recommendation model, and Respectively represent the loss functions for the next POI and POI category prediction, using cross entropy loss, For performance evaluation of access time prediction, mean square error (MSE) is used.

[0059] Compared with the prior art, the present invention has the following advantages:

[0060] The method of the present invention is suitable for recommending the next POI that the user is interested in. By constructing a regional user trajectory map, on the premise of filtering out the noise of personal unique preference, a social relationship network is modeled based on the correlation between different POIs, and the sign-in data information of all users is fully utilized. At the same time, the user information characteristics are deeply learned by using graph-based and sequence-based models, and the POI spatiotemporal-category information characteristics are mined by using relative visit time difference and absolute visit time. The sign-in data of the target user is comprehensively and comprehensively processed, utilized and adaptively integrated from multiple dimensions. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 It is a schematic diagram of a regional user trajectory map of the present invention;

[0062] Figure 2 is a schematic diagram of the present invention integrating the social relationship network and the spatiotemporal context information model;

[0063] Figure 3 is a schematic diagram of a gated graph convolutional neural network of the present invention;

[0064] Figure 4 is a schematic diagram of an encoder module of the present invention;

[0065] Figure 5 is a schematic diagram of the present invention in coping with a new user cold start test in the NYC data set;

[0066] Figure 6 It is a schematic diagram of the present invention in coping with the cold start test of new users in the TKY data set. DETAILED DESCRIPTION

[0067] In order to further illustrate the technical solution of the present invention, the present invention is further described below through embodiments.

[0068] A POI recommendation method integrating social relationship network and spatiotemporal context information in this embodiment includes the following steps:

[0069] Step 1: Obtain user sign-in sequence data, perform data preprocessing, and construct a regional user trajectory map;

[0070] The regional user trajectory diagram of this embodiment is as follows Figure 1 As shown, it is defined as: a directed graph consisting of a node set V, an edge set E, a node attribute S and a weight W. The node set V consists of the POIs that the user has visited historically. The edge set E is a set of directed edges. When a user visits two POIs p1 and p2 consecutively within a period of time, a directed edge e is added between p1 and p2. The weight W of each directed edge is equal to the number of times the sub-trajectory appears in the regional trajectory graph. The node attribute S contains the spatial coordinates, category and visit frequency of each POI.

[0071] Specifically, the FourSquare dataset is used as the dataset for the POI recommendation task that integrates social relationship networks and spatiotemporal context information in the present invention. The FourSquare-NYC dataset was collected in New York from April 2012 to February 2013, and the FourSquare-TKY dataset was collected in Tokyo during the same period. Each check-in record contains information such as user id, POI id, POI category, GPS coordinates and timestamp. The first 80% is taken as the training set, and the remaining 20% ​​is used as the test set. Delete POIs and users with less than 10 check-in records; divide the check-in sequence data of each user into tracks with an interval of 24 hours. If a track contains only one check-in record, delete the track data.

[0072] Step 2: Input the user check-in sequence data preprocessed in step 1 and the constructed domain user trajectory map into the pre-trained POI recommendation model (GSTC) to make the next POI recommendation.

[0073] The POI recommendation model of this embodiment is as follows: Figure 2 As shown in the figure, this model integrates social relationship network and spatiotemporal context information, including a social relationship learning module, a POI spatiotemporal-category feature fusion module, an encoder and a decoder.

[0074] Among them, the social relationship learning module: through the graph convolutional neural network integrated with the gated recurrent unit, the regional user trajectory graph is subjected to feature aggregation and noise filtering, the features of each POI are learned, and the feature vector E of each POI is obtained. p ;

[0075] like Figure 3 As shown in Figure 1, the specific process of the social relationship learning module includes:

[0076] Step 2.1: Calculate the normalized Laplacian matrix of the regional user trajectory graph adjacency matrix

[0077] Step 2.2: Use a single-layer graph convolutional neural network to normalize the Laplacian matrix Perform feature extraction;

[0078] Step 2.3: Update the hidden state of each node using the gated recurrent unit;

[0079] Step 2.4: Repeat steps 2.2 and 2.3 multiple times to obtain the feature vector h t The feature vector x extracted by the first single-layer graph convolutional neural network t (x1) is spliced ​​and fused to obtain the feature vector E of each POI p , the calculation process is expressed as follows:

[0080]

[0081] In the formula, H() represents a single-layer graph convolution operation, f T () represents a gated recurrent unit with T time steps, W G,C and b G,C denote weights and biases respectively, · denotes concatenation, is the Laplace matrix.

[0082] POI spatiotemporal-category feature fusion module: The time-space fusion module and the time-category fusion module are used to perform preliminary feature extraction and fusion on the POI spatiotemporal information of different dimensions in the user check-in sequence data, and the time-space fusion feature E of the user check-in activity is obtained. l,ts and time-category fusion feature E c,t ;

[0083] 1) Time-space fusion module: Adopt the encoding strategy of CTLE (Context and Time Aware Location Embeddings) model to obtain the timestamp feature vector E ts ; Input the coordinates of each POI into the fully connected layer to obtain the spatial feature vector E l ; The timestamp feature vector E ts and the spatial eigenvector E l The splicing is performed and input into the fully connected layer for feature fusion to obtain the time-space fusion feature E of the user's check-in activity. l,ts , the calculation process is expressed as follows:

[0084] E l,ts =σ(W l,ts [E ts ·E l ]+b l,ts ) (2)

[0085] Where W l,ts represents the learnable weight vector, b l,ts represents the deviation, σ represents the nonlinear activation function LeakyReLU;

[0086] 2) Time-category fusion module: Adopt the encoding strategy of the time encoding model time2vector to encode the absolute access time and obtain the time feature vector E t ; Embed each POI category into a category feature vector E c Represented; the category feature vector E c and the time feature vector E tThe splicing is then input into the fully connected layer for feature fusion to obtain the time-category fusion feature E of the user's check-in activity. c,t , the calculation process is expressed as follows:

[0087] E c,t =σ(W c,t [E c ·E t ]+b c,t ) (3)

[0088] Where W c,t represents the learnable weight vector, b c,t represents the bias, and σ represents the nonlinear activation function LeakyReLU.

[0089] Encoder: The user feature learning module and the spatiotemporal feature learning module are used to extract features of the user trajectory sequence and obtain the user feature Z u and the spatiotemporal characteristics Z t ;

[0090] The user trajectory sequence is defined as: the user's continuous check-in records within 48 hours are composed in chronological order; after obtaining the time-space fusion feature representation E of the user's check-in activity l,ts And the time-category fusion feature representation E c,t After that, the embedding representation of user check-in q is divided into two parts: and Each user's trajectory sequence Both are embedded by two user check-in sequences and express.

[0091] The structure of the encoder is as follows Figure 4 As shown, the specific process includes:

[0092] 1) User feature learning module: Construct the user feature vector E of the target user u’s personal information through an embedding layer u , construct the user check-in data into two modes: graph-based user trajectory graph and sequence-based user check-in sequence;

[0093] For the user trajectory graph, a graph convolutional neural network with a gated recurrent unit is used to extract features, and the obtained POI feature vector and user feature vector E u Splicing, input to the fully connected layer for nonlinear transformation, so as to obtain the graph-based POI-User fusion feature vector The calculation process is expressed as follows:

[0094]

[0095] in, and represent weight and bias respectively, represents splicing, σ L Represents the nonlinear activation function LeakyReLU;

[0096] For user check-in sequences, a lightweight self-attention mechanism and channel convolution are used to capture global context information and local context information respectively;

[0097] To capture global context information, firstly transform the feature vector E obtained by the social relationship learning module into p and user feature vector E u Splicing, output sequence-based POI-User fusion feature vector Then use a simplified lightweight self-attention mechanism to obtain global information, input it into the fully connected layer FC1 for nonlinear transformation, and obtain the sequence-based POI-User global feature vector The calculation process is expressed as follows:

[0098]

[0099] in, and Represent weights and biases respectively, Q, K, and V represent the input The same query weight matrix, key weight matrix and value weight matrix, d k represents the variance, T represents the transpose of the matrix, Represents the weight of the fully connected layer FC1. The encoder uses the softmax function to ensure that the sum of the attention weights is 1. At the same time, in order to avoid the overfitting problem, average pooling and layer normalization are used. AvgPooling represents average pooling, and LayerNorm represents layer normalization.

[0100] To capture local context information, the sequence-based POI-User fusion feature vector The two-dimensional channel convolution layer conv2D is input to obtain local information, and then input to the fully connected layer FC2 for nonlinear transformation to obtain the sequence-based POI-User local feature vector The calculation process is expressed as follows:

[0101]

[0102] in, Represents the weight of the fully connected layer FC2;

[0103] After obtaining the graph-based POI-User fusion feature vector Sequence-based POI-User global feature vector and the local eigenvector After that, the user feature learning module outputs the user feature Z u , defined by the weighted sum of these three feature vectors, in order to improve the adaptive feature fusion capability of the model, two aggregation hyperparameters α1 and α2 are introduced to reasonably allocate the fusion ratio. The calculation process is expressed as follows:

[0104]

[0105] Among them, σ t Represents the activation function tanh;

[0106] 2) Spatiotemporal feature learning module: given a user trajectory sequence Embedding representation of two user check-in sequences and Will and As input tensors, they are respectively passed into the spatiotemporal feature learning module to capture spatiotemporal features of different dimensions. The module has a multi-layer structure. For the pth layer, the input and Through the transformation of multi-head self-attention layers, the calculation process is expressed as follows:

[0107]

[0108] Among them, W Q , W K , W V represent the query weight matrix, key weight matrix, and value weight matrix respectively, E ij represents the correlation between the i-th and j-th check-ins, α ij Indicates the influence weight of other check-in activities on the current check-in. The encoder uses the softmax function to ensure that the sum of the attention weights is 1;

[0109] The output Head obtained by different attention heads i The images are stacked and the representations from different attention spaces are merged through linear transformation. In order to avoid overfitting and gradient disappearance problems, the encoder uses layer normalization LayerNorm and residual connection technology. The calculation process is expressed as follows:

[0110]

[0111] After the multi-head self-attention layer, the feedforward neural network FNN and activation function ReLU are used to give the model nonlinearity. The calculation process is expressed as follows:

[0112]

[0113] Among them, W1 and W2 represent weights, b1 and b2 represent biases, so the output of the encoder layer p is expressed as follows:

[0114]

[0115] After stacking the multi-head self-attention blocks multiple times, the spatiotemporal feature learning module outputs the spatiotemporal feature Z t , output features from time-space and time-category output features In order to improve the adaptive feature fusion capability of the model, two aggregation hyperparameters β1 and β2 are introduced to reasonably allocate the fusion ratio. The calculation process is expressed as follows:

[0116]

[0117] Among them, σ t represents the activation function tanh, p * Represents the last layer of the spatiotemporal feature learning module.

[0118] Decoder: User feature Z based on user context information u and the spatiotemporal characteristics Z t ,The next POI, visit time period and POI category are predicted respectively through three multi-layer perceptrons;

[0119] The decoder uses three multi-layer perceptrons to predict the next POI, visit time period, and POI category, respectively, as defined in the following formula:

[0120]

[0121] in, Indicates the next POI, Indicates the access time period. represents the POI category, W poi , W time , W cat Respectively represent their respective weights, b poi 、b time 、b cat Respectively represent their own deviations.

[0122] In addition, this method also includes the training of the POI recommendation model, and its loss function is defined as follows:

[0123]

[0124] In the formula, represents the overall loss function of the POI recommendation model, and Respectively represent the loss functions for the next POI and POI category prediction, using the common cross entropy loss, For performance evaluation of access time prediction, mean square error (MSE) is used.

[0125] Specifically, the Adam optimizer with an initial learning rate of 0.0001 is used to train the network, the batch size is set to 16, the network is trained for 150 epochs, and the results are obtained through multiple tests. See Tables 1 and 2 for details.

[0126] Table 1 Overall performance comparison of different methods on the NYC dataset

[0127]

[0128] Table 2 Overall performance comparison of different methods on the TKY dataset

[0129]

[0130] Accuracy Acc@k and average reciprocal ranking MRR are used as measurement indicators on the FourSquare dataset. Tables 1 and 2 compare the measurement indicators of the method of the present invention and other methods for POI recommendation on the NYC and TKY datasets. It can be seen from Tables 1 and 2 that the method of the present invention is superior to other methods in terms of accuracy Acc@k and average reciprocal ranking MRR. Compared with GETNext, the method of the present invention improves Acc@1 by 2.02%, Acc@5 by 2.99%, Acc@10 by 2.32%, Acc@20 by 2.81%, and MRR by 2.34% in the NYC dataset; in the TKY dataset, Acc@1 improves by 0.85%, Acc@5 by 2.02%, Acc@10 by 2.68%, Acc@20 by 5.22%, and MRR by 1.38%.

[0131] The effectiveness of the POI recommendation network GSTC of the present invention is verified by changing the input user trajectory length n. Figure 5 To investigate the impact of user trajectory length n on Acc@10 in the NYC dataset, Figure 6The impact of the input user trajectory length n on Acc@10 in the TKY dataset. In order to test the model's performance in dealing with new user cold start, the value of n is set to 2 to 12 to simulate different sparsity levels of user check-in data while keeping other parameters unchanged. The trend in the figure shows that when n=2, the accuracy of the GSTC model on the NYC and TKY datasets is significantly better than that of the baseline model, and the larger the dataset, the more obvious the leading advantage. As the input trajectory length continues to increase, the recommendation effect of the existing model shows an overall upward trend. This shows that the GSTC model can effectively utilize the POI-POI social relationship network modeled by the gated graph convolutional neural network to provide more accurate recommendations for target users, thereby alleviating the cold start problem of new users.

[0132] The above shows and describes the main features and advantages of the present invention. It is obvious to those skilled in the art that the present invention is not limited to the details of the above exemplary embodiments, and that the present invention can be implemented in other specific forms without departing from the spirit or essential features of the present invention. Therefore, no matter from which point of view, the embodiments should be regarded as exemplary and non-restrictive. The scope of the present invention is defined by the appended claims rather than the above description, and it is intended that all changes that fall within the meaning and scope of the equivalent elements of the claims are included in the present invention.

[0133] In addition, it should be understood that although the present specification is described according to implementation modes, not every implementation mode contains only one independent technical solution. This description of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment may also be appropriately combined to form other implementation modes that can be understood by those skilled in the art.

Claims

1. A POI recommendation method integrating social relationship network and spatiotemporal context information, characterized in that: The following steps are involved: Step 1: Obtain user sign-in sequence data, perform data preprocessing, and construct a regional user trajectory map; Step 2: Input the user check-in sequence data preprocessed in step 1 and the constructed domain user trajectory map into the pre-trained POI recommendation model to make the next POI recommendation; the POI recommendation model integrates the social relationship network and spatiotemporal context information, including a social relationship learning module, a POI spatiotemporal-category feature fusion module, an encoder and a decoder; the social relationship learning module performs feature aggregation and noise filtering on the regional user trajectory map through a graph convolutional neural network integrated with a gated recurrent unit, learns the features of each POI, and obtains the feature vector E of each POI. p The POI spatiotemporal-category feature fusion module performs preliminary feature extraction and fusion on the POI spatiotemporal information of different dimensions in the user check-in sequence data through the time-space fusion module and the time-category fusion module to obtain the time-space fusion feature E of the user check-in activity. l,ts and time-category fusion feature E c,t The encoder extracts features from the user trajectory sequence through the user feature learning module and the spatiotemporal feature learning module to obtain the user feature Z u and the spatiotemporal characteristics Z t ; The decoder is based on user feature Z u and the spatiotemporal characteristics Z t ,The next POI, the visiting time period and the POI category are predicted respectively through three multi-layer perceptrons.

2. A POI recommendation method integrating social relationship network and spatiotemporal context information according to claim 1, characterized in that: The regional user trajectory graph in step 1 is defined as: a directed graph consisting of a node set V, an edge set E, a node attribute S and a weight W, wherein the node set V consists of the POIs that the user has visited historically, and the edge set E is a set of directed edges. When a user visits two POIs p1 and p2 consecutively within a period of time, a directed edge e is added between p1 and p2. The weight W of each directed edge is equal to the number of times the sub-trajectory appears in the regional trajectory graph, and the node attribute S includes the spatial coordinates, category and visit frequency of each POI.

3. The POI recommendation method integrating social relationship network and spatiotemporal context information according to claim 1, characterized in that: The specific process of the social relationship learning module includes: Step 2.1: Calculate the normalized Laplacian matrix of the regional user trajectory graph adjacency matrix Step 2.2: Use a single-layer graph convolutional neural network to normalize the Laplacian matrix Perform feature extraction; Step 2.3: Update the hidden state of each node using a gated recurrent unit. Step 2.4: After repeating steps 2.2 and 2.3 multiple times, the obtained feature vector is compared with the feature vector x extracted by the first single-layer graph convolutional neural network. t Splice and merge to get the feature vector E of each POI p , the calculation process is expressed as follows: In the formula, H() represents a single-layer graph convolution operation, f T () represents a gated recurrent unit with T time steps, W G,C and b G,C denote weights and biases respectively, · denotes concatenation, is the Laplace matrix.

4. The POI recommendation method integrating social relationship network and spatiotemporal context information according to claim 3, characterized in that: The specific process of the POI spatiotemporal-category feature fusion module includes: 1) Time-space fusion module: Adopt the encoding strategy of the CTLE model to obtain the timestamp feature vector E ts ; Input the coordinates of each POI into the fully connected layer to obtain the spatial feature vector E l ; The timestamp feature vector E ts and the spatial eigenvector E l The splicing is performed and input into the fully connected layer for feature fusion to obtain the time-space fusion feature E of the user's check-in activity. l,ts , the calculation process is expressed as follows: E l,ts σ(W l,ts [E ts ·E l ]+b l,ts ) (2) Where W l,ts represents the learnable weight vector, b l,ts represents the deviation, σ represents the nonlinear activation function Leaky ReLU; 2) Time-category fusion module: Adopt the encoding strategy of the time encoding model time2vector to encode the absolute access time and obtain the time feature vector E t ; Embed each POI category into a category feature vector E c Represented; the category feature vector E c and the time feature vector E t The splicing is then input into the fully connected layer for feature fusion to obtain the time-category fusion feature E of the user's check-in activity. c,t , the calculation process is expressed as follows: E c,t σ(W c,t [E c ·E t ]+b c,t ) (3) Where W c,t represents the learnable weight vector, b c,t represents the bias, and σ represents the nonlinear activation function Leaky ReLU.

5. The POI recommendation method integrating social relationship network and spatiotemporal context information according to claim 4, characterized in that: The user trajectory sequence is defined as: consisting of the user's continuous check-in records in chronological order; after obtaining the time-space fusion feature representation E of the user's check-in activity l,ts And the time-category fusion feature representation E c,t After that, the embedding representation of user check-in q is divided into two parts: and Each user's trajectory sequence Both are embedded by two user check-in sequences and express.

6. The POI recommendation method integrating social relationship network and spatiotemporal context information according to claim 5, characterized in that: The specific process of the encoder includes: 1) User feature learning module: Construct the user feature vector E of the target user u’s personal information through an embedding layer u , construct the user check-in data into two modes: graph-based user trajectory graph and sequence-based user check-in sequence; For the user trajectory graph, a graph convolutional neural network with a gated recurrent unit is used to extract features, and the obtained POI feature vector and user feature vector E u Splicing, input to the fully connected layer for nonlinear transformation, so as to obtain the graph-based POI-User fusion feature vector The calculation process is expressed as follows: in, and represent weight and bias respectively, represents splicing, σ L Represents the nonlinear activation function LeakyReLU; For user check-in sequences, a lightweight self-attention mechanism and channel convolution are used to capture global context information and local context information respectively; To capture global context information, firstly transform the feature vector E obtained by the social relationship learning module into p and user feature vector E u Splicing, output sequence-based POI-User fusion feature vector Then use a simplified lightweight self-attention mechanism to obtain global information, input it into the fully connected layer FC1 for nonlinear transformation, and obtain the sequence-based POI-User global feature vector The calculation process is expressed as follows: in, and Represent weights and biases respectively, Q, K, and V represent the input The same query weight matrix, key weight matrix and value weight matrix, d k represents the variance, T represents the transpose of the matrix, Represents the weight of the fully connected layer FC1, AvgPooling represents average pooling, and LayerNorm represents layer normalization; To capture local context information, the sequence-based POI-User fusion feature vector The local information is input into the two-dimensional channel convolution layer conv2D to obtain local information, and then input into the fully connected layer FC2 for nonlinear transformation to obtain the sequence-based POI-User local feature vector The calculation process is expressed as follows: in, Represents the weight of the fully connected layer FC2; After obtaining the graph-based POI-User fusion feature vector Sequence-based POI-User global feature vector and the local eigenvector After that, the user feature learning module outputs the user feature Z u , which is defined by the weighted sum of these three feature vectors, and two aggregation hyperparameters α1 and α2 are introduced to reasonably allocate the fusion ratio. The calculation process is expressed as follows: Among them, σ t represents the activation function tanh; 2) Spatiotemporal feature learning module: given a user trajectory sequence Embedding representation of two user check-in sequences and Will and As input tensors, they are respectively passed into the spatiotemporal feature learning module to capture spatiotemporal features of different dimensions. The module has a multi-layer structure. For the pth layer, the input and Through the transformation of multi-head self-attention layers, the calculation process is expressed as follows: Among them, W Q , W K , W V Represent the query weight matrix, key weight matrix and value weight matrix respectively, E ij represents the correlation between the i-th and j-th check-ins, α ij Indicates the influence weight of other check-in activities on the current check-in; The output Head obtained by different attention heads i The layers are stacked and the representations from different attention spaces are merged through linear transformation. The encoder uses layer normalization LayerNorm and residual connection technology. The calculation process is expressed as follows: After the multi-head self-attention layer, the feedforward neural network FNN and activation function ReLU are used to give the model nonlinearity. The calculation process is expressed as follows: Among them, W1 and W2 represent weights, b1 and b2 represent biases, so the output of the encoder layer p is expressed as follows: After stacking the multi-head self-attention blocks multiple times, the spatiotemporal feature learning module outputs the spatiotemporal feature Z t , output features from time-space and time-category output features The weighted sum definition of , introduces two aggregation hyperparameters β1 and β2 to reasonably distribute the fusion ratio. The calculation process is expressed as follows: Among them, σ t represents the activation function tanh, p * Represents the last layer of the spatiotemporal feature learning module.

7. The POI recommendation method integrating social relationship network and spatiotemporal context information according to claim 1, characterized in that: The decoder uses three multi-layer perceptrons to predict the next POI, visit time period and POI category respectively, which are defined as follows: in, Indicates the next POI, Indicates the access time period. represents the POI category, W poi , W time , W cat Respectively represent their respective weights, b poi , b time , b cat Respectively represent their own deviations.

8. A POI recommendation method integrating social relationship network and spatiotemporal context information according to any one of claims 1 to 7, characterized in that: It also includes the training of the POI recommendation model, and its loss function is defined as follows: In the formula, represents the overall loss function of the POI recommendation model, and Respectively represent the loss functions for the next POI and POI category prediction, using cross entropy loss, For performance evaluation of access time prediction, mean square error (MSE) is used.

Images