Method for aligning cross-social network user identities based on user spatio-temporal data

Abstract

The application provides a cross-social network user identity alignment method based on user space-time data, which has the following characteristics and comprises the following steps: step S1, pre-processing user check-in data A and user check-in data B respectively to obtain user trajectory T A and user trajectory T β respectively; step S2, inputting user trajectory T A and user trajectory T B into a user identity alignment prediction model to obtain a user identity alignment result, wherein the user identity alignment prediction model comprises a vectorization module, a trajectory cross-attention module, a Bi-LSTM module, a local attention module and an output module. In summary, the method can obtain a more accurate user identity alignment result.

Description

Technical Field

[0001] This invention relates to the field of online social networks, and more specifically to a method for aligning user identities across social networks based on user spatiotemporal data. Background Technology

[0002] With the widespread adoption of mobile devices, online social networks have become increasingly popular, with various types of social platforms emerging and becoming an important part of people's daily lives. Due to the diversity of social platform functions, people often join multiple platforms simultaneously to meet different needs, such as Douban, Zhihu, and Sina Weibo. Users leave various types of personalized information on different social platforms, such as personal information, pictures, and check-in locations. Integrating this information from different social platforms allows for a more comprehensive user profile and personalized recommendations. Therefore, cross-social network user identification has gained widespread attention in recent years.

[0003] Cross-social network user identity alignment matches user identities from different platforms, connecting accounts belonging to the same individual to provide a more comprehensive understanding of a user's social behavior and facilitate research on social network user behavior. Regarding cross-social network user identity alignment, some researchers have conducted studies based on single-type user-generated data, such as user attributes, user-generated content, and network structure. However, due to errors in user attribute information during registration, the sparsity of social network structures, and the short-text nature of user-generated content, researchers find it difficult to extract sufficient user features for alignment.

[0004] The geospatial information of user check-ins is relatively ideal matching data. Currently, end-to-end methods are used to learn representations of check-in locations and movement trajectories to predict user alignment results. However, these methods still have certain problems to solve. On the one hand, due to the differences between social platforms, users' check-ins at the same location may have some deviation. Traditional location representation methods, such as one-hot encoding, have a semantic gap problem and cannot learn the potential semantic relationships between locations. On the other hand, users have certain preferences for check-in behavior on different platforms, resulting in a large imbalance in user check-in data across different platforms. This greatly interferes with accurately representing user behavior patterns.

[0005] In summary, there is still significant room for improvement in existing technologies for aligning user identities across social networks using user check-in data. Summary of the Invention

[0006] This invention is made to solve the above-mentioned problems, and aims to provide a method for aligning user identities across social networks based on user spatiotemporal data.

[0007] This invention provides a cross-social network user identity alignment method based on user spatiotemporal data, used to align user identity across platforms A based on user u i User check-in data A and user u in platform B j From user check-in data B, we obtain the user identity alignment results for user A and user B, which have the following characteristics and include the following steps: Step S1, preprocess user check-in data A and user check-in data B respectively to obtain user trajectory T. A and user trajectory T B Step S2, transfer user trajectory T A and user trajectory T B The user identity alignment prediction model is input together to obtain the user identity alignment result. The user identity alignment prediction model includes a vectorization module, a trajectory cross-attention module, a Bi-LSTM module, a local attention module, and an output module. The vectorization module is used to convert the user trajectory T into its corresponding data. A and user trajectory T B Vectorization is performed to obtain the check-in location vector X. i and the sign-in location vector X j The trajectory cross-attention module is used to process the check-in location vector X according to the trajectory cross-attention mechanism. i and the sign-in location vector X j The enhancement is performed to obtain the enhanced check-in location vector X′. i and the enhanced check-in location vector X′ j The Bi-LSTM module contains a bidirectional long short-term memory (LSTM) mental network, used to enhance the check-in location vector X′. i and the enhanced check-in location vector X′ j Feature extraction is performed to obtain the trajectory hiding features h. i Trajectory Hiding Features h j The local attention module is used to hide features h based on the trajectory. j Enhance the sign-in location vector X′ i Trajectory Hiding Features h i The final embedding representation vector s′ is obtained. i Based on trajectory hiding features h i Enhance the check-in location vector X′ j Trajectory Hiding Features h j The final embedding representation vector s′ is obtained. j The output module is used to determine the final embedded representation vector s′. i and the final embedding representation vector s′ j Get user trajectory T A and user trajectory T B Matching probability As a result of user identity alignment.

[0008] The cross-social network user identity alignment method based on user spatiotemporal data provided by the present invention may also have the following features: wherein the check-in data includes GPS information and corresponding timestamp information of each check-in location, and in step S1, the preprocessing is: sorting the timestamp information and corresponding GPS information in the user check-in data according to the time order to obtain the user trajectory.

[0009] The cross-social network user identity alignment method based on user spatiotemporal data provided by this invention may also have the following features: the vectorization module obtains the corresponding embedded vector representation for each check-in position in the input user trajectory through a one-hot encoding algorithm, and then maps all the embedded vector representations to a low-dimensional vector space and performs a concatenation operation to obtain the check-in position vector corresponding to the user trajectory.

[0010] The cross-social network user identity alignment method based on user spatiotemporal data provided in this invention may also have the following feature: wherein, the enhanced check-in location vector X′ i The calculation formula is: X′ i =softmax(Υ) ij V j +X i , In the formula V j Let X be the sign-in location vector. j Location features mapped to the V feature space, softmax() is a non-linear activation function, Q i Let X be the sign-in location vector. i Location features mapped to the Q feature space, K j Let X be the sign-in location vector. j The location features are mapped to the K feature space, ⊙ represents the element-wise multiplication operation of the matrix, λ is the spatiotemporal correlation balance parameter, and distance() is the semi-sine distance formula. for Chinese user u i The corresponding m-th check-in position and user u j Spatial correlation of the corresponding nth check-in location For user trajectory T A Chinese user u i The GPS information for the corresponding m-th check-in location. For user trajectory T B Chinese user u j The GPS information for the corresponding nth check-in location, where α is the distance adjustment parameter. for Chinese user u i The corresponding m-th check-in position and user uj The time relevance of the corresponding nth check-in location For user trajectory T A Chinese user u i The timestamp information corresponding to the m-th check-in location. For user trajectory T B Chinese user u j The timestamp information of the corresponding nth check-in location, β is the time adjustment parameter, || is the first normal form operation, e is a constant, and the check-in location vector X′ is enhanced. j The calculation formula is: X′ j =softmax(γ) ji V i +X j , In the formula V i Let X be the sign-in location vector. i Location features mapped to the V feature space, Q j Let X be the sign-in location vector. j Location features mapped to the Q feature space, K i Let X be the sign-in location vector. i Location features mapped to the K feature space.

[0011] The cross-social network user identity alignment method based on user spatiotemporal data provided in this invention may also have the following feature: wherein the final embedded representation vector s′ i The calculation formula is: s′ i =σ(W T [s i h i ]+b T ), a m，j =w T (tanh([W l X′ i，m W g h j In the formula, σ is the activation function, and W T For training parameters, b T For the bias term, l i For user trajectory T A The length of x′ i，m To enhance the check-in location vector X′ i User trajectory T A The vector corresponding to the m-th check-in position, tanh() is the activation function, W l W g Both and w are training parameters, and the final embedding representation vector s′ j The calculation formula is: s′j =σ(W T [s j h j ]+b T ), a m，i =w T (tanh([W l X′ j ,m;W g h i ])), where l j For user trajectory T B The length of x′ j，m To enhance the check-in location vector X′ j User trajectory T B The vector corresponding to the m-th check-in position.

[0012] The cross-social network user identity alignment method based on user spatiotemporal data provided in this invention may also have the following feature: wherein, the matching probability The calculation formula is: In the formula, W1 and W2 are both training parameter matrices, and b1 and b2 are both bias terms.

[0013] The cross-social network user identity alignment method based on user spatiotemporal data provided in this invention may also have the following feature: The user identity alignment prediction model is trained using a training dataset constructed from existing social network data and a loss function Loss. The training dataset includes M sample pairs, each sample pair including a sample and its corresponding real label. The formula for calculating the loss function Loss is: In the formula Let y be the matching probability obtained after inputting the user identity alignment prediction model into the i-th sample. i Let be the true label corresponding to the i-th sample.

[0014] The role and effect of invention

[0015] The cross-social network user identity alignment method based on user spatiotemporal data according to the present invention improves the accuracy of matching probability prediction, i.e., the prediction accuracy of user identity alignment results, by enhancing the check-in location vector through a trajectory cross-attention mechanism to obtain an enhanced check-in location vector that better expresses the spatiotemporal relationship between locations, and by extracting key location features through a local attention mechanism to obtain a final embedded representation vector expressing user behavior features. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating the cross-social network user identity alignment method based on user spatiotemporal data in an embodiment of the present invention.

[0017] Figure 2 This is a schematic diagram of the framework of the user identity alignment prediction model in an embodiment of the present invention;

[0018] Figure 3 This is a schematic diagram illustrating the evaluation results of various methods on the Foursquare-Twitter dataset in an embodiment of the present invention;

[0019] Figure 4 This is a schematic diagram illustrating the evaluation results of various methods on the Instagram-Twitter dataset in an embodiment of the present invention;

[0020] Figure 5 This is a schematic diagram illustrating the evaluation results of each model on the Foursquare-Twitter dataset in an embodiment of the present invention;

[0021] Figure 6 This is a schematic diagram illustrating the evaluation results of each model on the Instagram-Twitter dataset in an embodiment of the present invention. Detailed Implementation

[0022] To make the technical means, creative features, objectives and effects of this invention easy to understand, the following embodiments, in conjunction with the accompanying drawings, specifically illustrate the cross-social network user identity alignment method based on user spatiotemporal data of this invention.

[0023] The cross-social network user identity alignment method based on user spatiotemporal data in this embodiment is used to align user identity based on user u in platform A. i User check-in data A and user u in platform B j From user check-in data B, we obtain the user identity alignment results for user A and user B.

[0024] Figure 1 This is a flowchart illustrating the cross-social network user identity alignment method based on user spatiotemporal data in an embodiment of the present invention.

[0025] like Figure 1 As shown, the cross-social network user identity alignment method based on user spatiotemporal data in this embodiment includes the following steps:

[0026] Step S1: Preprocess user check-in data A and user check-in data B respectively to obtain user trajectory T. A and user trajectory T B .

[0027] The check-in data includes GPS information and corresponding timestamp information for each check-in location. The preprocessing involves sorting the timestamp information and corresponding GPS information in the user check-in data according to time order to obtain the user trajectory.

[0028] Step S2, record the user trajectory T A and user trajectory T B Input them into the user identity alignment prediction model to obtain the user identity alignment results.

[0029] Figure 2 This is a schematic diagram of the framework of the user identity alignment prediction model in an embodiment of the present invention.

[0030] like Figure 2 As shown, the user identity alignment prediction model 100 includes a vectorization module 10, a trajectory cross-attention module 20, a Bi-LSTM module 30, a local attention module 40, and an output module 50.

[0031] Vectorization module 10 is used to convert user trajectory T into vectorization module T. A and user trajectory T B Vectorization is performed to obtain the check-in location vector X. i and the sign-in location vector X j .

[0032] In this embodiment, the vectorization module 10 obtains the corresponding embedded vector representation for each check-in location, i.e., GPS information and timestamp information in the input user trajectory, through a one-hot encoding algorithm. Then, it maps all the embedded vector representations to a low-dimensional vector space and performs a concatenation operation to obtain the check-in location vector corresponding to the user trajectory. In this embodiment, dimensionality reduction is achieved through linear transformation so that the embedded vector representation is mapped to a low-dimensional vector space.

[0033] In this embodiment, the one-hot encoding algorithm works as follows: A vocabulary of check-in locations with size 'a' and a vocabulary of check-in times with size 'b' are introduced. This ensures that the GPS information and timestamp information for each check-in location are 1 in their respective vocabulary, and 0 in the remaining positions. This yields 0-1 vectors of dimensions 'a' and 'b', which are then used as embedding vectors. Finally, a concatenation operation transforms each user trajectory into a d-dimensional vector matrix, i.e., the check-in location vector.

[0034] The trajectory cross-attention module 20 is used to process the check-in position vector X according to the trajectory cross-attention mechanism. i and the sign-in location vector X j The enhancement is performed to obtain the enhanced check-in location vector X′. i and the enhanced check-in location vector X′ j .

[0035] Among them, the enhanced check-in location vector X′ jThe calculation formula is:

[0036] X′ i =softmax(Υ) ij V j +X i ,

[0037]

[0038]

[0039]

[0040] In the formula V j Let X be the sign-in location vector. j Location features mapped to the V feature space, softmax() is a non-linear activation function, Q i Let X be the sign-in location vector. i Location features mapped to the Q feature space, K j Let X be the sign-in location vector. j The location features are mapped to the K feature space, ⊙ represents the element-wise multiplication operation of the matrix, λ is the spatiotemporal correlation balance parameter, and distance() is the semi-sine distance formula. for Chinese user u i The corresponding m-th check-in position and user u j Spatial correlation of the corresponding nth check-in location For user trajectory T A Chinese user u i The GPS information for the corresponding m-th check-in location. For user trajectory T B Chinese user u j The GPS information for the corresponding nth check-in location, where α is the distance adjustment parameter. for Chinese user u i The corresponding m-th check-in position and user u j The time relevance of the corresponding nth check-in location For user trajectory T A Chinese user u i The timestamp information corresponding to the m-th check-in location. For user trajectory T B Chinese user u j The timestamp information of the corresponding nth check-in location, β is the time adjustment parameter, || is the first normal form operation, and e is a constant. In this embodiment... This represents the spatial correlation between two check-in locations on different trajectories. The temporal correlation between two check-in locations in different trajectories.

[0041] Enhance the check-in location vector X′ j The calculation formula is:

[0042] X′ j =softmax(Υ) ji V i +X j ,

[0043]

[0044] In the formula V i Let X be the sign-in location vector. i Location features mapped to the V feature space, Q j Let X be the sign-in location vector. j Location features mapped to the Q feature space, K i Let X be the sign-in location vector. i Location features mapped to the K feature space.

[0045] In this embodiment, the check-in location vector is mapped to the Q, K, V feature space through the trajectory cross-attention module 20. By using the trajectory cross-attention mechanism that integrates the spatiotemporal correlation of location, the interdependence with the target location is introduced, thereby obtaining an enhanced check-in location vector.

[0046] The Bi-LSTM module 30 contains a bidirectional long short-term memory (LSTM) mental network used to enhance the check-in location vector X′. i and the enhanced check-in location vector X′ j Feature extraction is performed to obtain the trajectory hiding features h. i Trajectory Hiding Features h j .

[0047] In this embodiment, the trajectory hiding feature h i Trajectory Hiding Features h j The calculation formulas are as follows:

[0048] h i =Bi-LSTM(X′) i ),

[0049] h j =Bi-LSTM(X) j ),

[0050] In the formula, Bi-LSTM() is a bidirectional long short-term memory mental network.

[0051] Local attention module 40 is used to hide features h based on the trajectory. i Enhance the check-in location vector X′ iTrajectory Hiding Features h i The final embedding representation vector s′ is obtained. i Based on trajectory hiding features h i Enhance the check-in location vector X′ j Trajectory Hiding Features h j The final embedding representation vector s′ is obtained. j .

[0052] Wherein, the final embedding representation vector s′ j The calculation formula is:

[0053] s′ i =σ(W T [s i h i ]+b T ),

[0054]

[0055] a m，j =w T (tanh([W l X′ i,m W g h j ])),

[0056] In the formula, σ is the activation function, and W T For training parameters, b T For the bias term, l i For user trajectory Y A The length of x′ i，m To enhance the check-in location vector X′ i User trajectory T A The vector corresponding to the m-th check-in position, tanh() is the activation function, W l W g Both and w are training parameters.

[0057] The final embedding representation vector s′ j The calculation formula is:

[0058] s′ j =σ(W T [s j h j ]+b T ),

[0059]

[0060] a m,i =w T (tanh([W l X′ j，m Wg h i ])),

[0061] In the formula l j For user trajectory T B The length of x′ j，m To enhance the check-in location vector X′ j User trajectory T B The vector corresponding to the m-th check-in position.

[0062] In this embodiment, the local attention module 40 extracts key sign-in location features s through a local attention mechanism. i and s j Then, the key check-in location features are integrated with the corresponding trajectory hiding features to obtain the final embedding representation vector.

[0063] Output module 50 is used to determine the final embedded representation vector s′ i and the final embedding representation vector s′ j Get user trajectory T A and user trajectory T B Matching probability As a result of user identity alignment.

[0064] Among them, the matching probability The calculation formula is:

[0065]

[0066] In the formula, W1 and W2 are both training parameter matrices, and b1 and b2 are both bias terms.

[0067] In this embodiment, the output module 50 includes a multilayer perceptron with a hidden layer to simulate nonlinear changes, integrate features of different trajectories, and then use the sigmoid() function to obtain the matching probability. Dropout is provided between layers in the output module 50 to prevent overfitting.

[0068] In this embodiment, a training dataset is constructed using existing social network data, and the user identity alignment prediction model 100 is trained using the loss function Loss until the model converges or reaches the maximum number of iterations, thus obtaining the trained user identity alignment prediction model 100.

[0069] The training dataset consists of M sample pairs, each pair including a sample and its corresponding true label. In this example, M / 2 sample pairs are positive sample pairs, and M / 2 sample pairs are negative sample pairs. The samples in the positive sample pairs are positive samples (T). A+ T B+ ), L A+ Y represents the trajectory sequence of users in platform A.B+ For T A+ The corresponding user trajectory sequence on platform B, the samples in the positive sample pairs are positive samples (T) A+ T B- ), T B- For non-T A+ The corresponding user in platform B is related to T. A+ A sequence of trajectories with the same check-in location.

[0070] The formula for calculating the loss function (Loss) is as follows:

[0071]

[0072] In the formula Let y be the matching probability obtained after inputting the user identity alignment prediction model into the i-th sample. i Let be the true label corresponding to the i-th sample.

[0073] In this embodiment, the cross-social network user identity alignment method (ours method) based on user spatiotemporal data of the present invention, as well as the existing POIS method, STUL method, GKR-KDE method and DPLINK* method, are evaluated using the existing Foursquare-Twitter dataset and Instagram-Twitter dataset. The evaluation metrics include precision (pre), recall, F1 score (f1), and AUC (auc). The evaluation settings include a mini-batch size of 64, a learning rate of 0.005, and the Adam optimizer.

[0074] Figure 3 This is a schematic diagram illustrating the evaluation results of various methods on the Foursquare-Twitter dataset in an embodiment of the present invention.

[0075] Figure 4 This is a schematic diagram illustrating the evaluation results of various methods on the Instagram-Twitter dataset in an embodiment of the present invention.

[0076] like Figure 3 , Figure 4 As shown, the horizontal axis represents each method and its corresponding evaluation index, and the vertical axis represents the value of the evaluation index. It can be seen that on the two existing datasets, the cross-social network user identity alignment method based on user spatiotemporal data of the present invention can achieve the best index evaluation results compared with other existing methods. Therefore, the cross-social network user identity alignment method based on user spatiotemporal data of the present invention can have more accurate prediction results, i.e., user identity alignment results.

[0077] In this embodiment, the user identity alignment prediction model 100 without the trajectory cross attention module 20 is used as the w / o-CTA model, and the user identity alignment prediction model 100 without the local attention module 40 is used as the w / o-LA model. The user identity alignment prediction model 100 of this embodiment, i.e., the Ours model, is evaluated on the Foursquare-Twitter dataset and the Instagram-Twitter dataset, respectively.

[0078] Figure 5 This is a schematic diagram illustrating the evaluation results of each model on the Foursquare-Twitter dataset in an embodiment of the present invention.

[0079] Figure 6 This is a schematic diagram illustrating the evaluation results of each model on the Instagram-Twitter dataset in an embodiment of the present invention.

[0080] like Figure 5 , Figure 6 As shown, the horizontal axis represents each evaluation metric and its corresponding model, while the vertical axis represents the numerical value of the evaluation metric. It can be seen that the recall of the w / o-CTA model is significantly reduced in both datasets. Therefore, the cross-attention mechanism can help to better express the spatiotemporal relationship between locations and is helpful in solving the heterogeneity problem of cross-social network check-in data. The w / o-LA model shows a decrease in precision in both datasets. Since the user's action range is generally limited, extracting key location features has a more positive effect on expressing user behavior features. That is, the local attention mechanism can significantly improve the method's results in predicting user alignment.

[0081] The role and effect of the embodiments

[0082] According to the cross-social network user identity alignment method based on user spatiotemporal data involved in this embodiment, on the one hand, the check-in location vector is enhanced through a trajectory cross-attention mechanism to obtain an enhanced check-in location vector that better expresses the spatiotemporal relationship between locations; on the other hand, key location features are extracted through a local attention mechanism to obtain the final embedded representation vector that expresses user behavior features, thereby improving the prediction accuracy of the matching probability, i.e., the user identity alignment result. In summary, this method can obtain more accurate user identity alignment results.

[0083] The above embodiments are preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention.

Claims

1. A method for aligning user identities across social networks based on user spatiotemporal data, used to align user identities based on user u in platform A. i User check-in data A and user u in platform B j From user check-in data B, the user identity alignment results of user A and user B are obtained, characterized in that... Includes the following steps: Step S1: Preprocess the user check-in data A and the user check-in data B respectively to obtain the user trajectory T. A and user trajectory T B ; Step S2, the user trajectory T A and the user trajectory T B The user identity alignment prediction model is input together to obtain the user identity alignment result. The user identity alignment prediction model includes a vectorization module, a trajectory cross-attention module, a Bi-LSTM module, a local attention module, and an output module. The vectorization module is used to convert the user trajectory T into... A and the user trajectory T B Vectorization is performed to obtain the check-in location vector X. i and the sign-in location vector X j , The trajectory cross-attention module is used to process the check-in location vector X according to the trajectory cross-attention mechanism. i and the check-in location vector X j The enhancement is performed to obtain the enhanced check-in location vector X′. i and the enhanced check-in location vector X′ j , The Bi-LSTM module includes a bidirectional long short-term memory neural network, used to process the enhanced check-in location vector X′ respectively. i and the enhanced check-in location vector X′ j Feature extraction is performed to obtain the trajectory hiding features h. i Trajectory Hiding Features h j , The local attention module is used to hide the trajectory feature h. j The enhanced check-in location vector X′ i and the trajectory hiding feature h i The final embedding representation vector s′ is obtained. i According to the trajectory hiding feature h i The enhanced check-in location vector X′ j and the trajectory hiding feature h j The final embedding representation vector s′ is obtained. j , The output module is used to determine the final embedding representation vector s′ based on the output module. i and the final embedding representation vector s′ j The user trajectory T is obtained A and the user trajectory T B Matching probability This serves as the result of the user identity alignment.

2. The cross-social network user identity alignment method based on user spatiotemporal data according to claim 1, characterized in that: in, The check-in data includes GPS information and corresponding timestamp information for each check-in location. In step S1, the preprocessing is as follows: sorting the timestamp information and the corresponding GPS information in the user check-in data according to time order to obtain the user trajectory.

3. The cross-social network user identity alignment method based on user spatiotemporal data according to claim 1, characterized in that: in, The vectorization module obtains the corresponding embedded vector representation for each check-in position in the input user trajectory using a one-hot encoding algorithm, then maps all the embedded vector representations to a low-dimensional vector space and performs a concatenation operation to obtain the check-in position vector corresponding to the user trajectory.

4. The cross-social network user identity alignment method based on user spatiotemporal data according to claim 1, characterized in that: in, The enhanced check-in location vector X′ i The calculation formula is: X′ i =softmax(γ ij )V j +X i , In the formula V j Let X be the check-in location vector. j Location features mapped to the V feature space, softmax() is a non-linear activation function, Q i Let X be the check-in location vector. i Location features mapped to the Q feature space, K j Let X be the check-in location vector. j The location features are mapped to the K feature space, ⊙ represents the element-wise multiplication operation of the matrix, λ is the spatiotemporal correlation balance parameter, and distance() is the semi-sine distance formula. for Chinese user u i The corresponding m-th check-in position and user u j Spatial correlation of the corresponding nth check-in location For user trajectory T A Chinese user u i The GPS information for the corresponding m-th check-in location. For user trajectory T B Chinese user u j The GPS information for the corresponding nth check-in location, where α is the distance adjustment parameter. for Chinese user u i The corresponding m-th check-in position and user u j The time relevance of the corresponding nth check-in location For user trajectory T A Chinese user u i The timestamp information corresponding to the m-th check-in location. For user trajectory T B Chinese user u j The timestamp information for the corresponding nth check-in location, β is the time adjustment parameter, || is the first-order normal form operation, and e is a constant. The enhanced check-in location vector X′ j The calculation formula is: X′ j =softmax(Υ ji )V i +X j , In the formula V i Let X be the check-in location vector. i Location features mapped to the V feature space, Q j Let X be the check-in location vector. j Location features mapped to the Q feature space, K i Let X be the check-in location vector. i Location features mapped to the K feature space.

5. The cross-social network user identity alignment method based on user spatiotemporal data according to claim 1, characterized in that: in, The final embedded representation vector s′ i The calculation formula is: s′ i =σ(W T [s i ;h i ]+b T ), a m，j =w T (tanh([W l X′ i ,m;W g h j ])), In the formula, σ is the activation function, and W T For training parameters, b T For the bias term, l i For user trajectory T A The length of x′ i，m To enhance the check-in location vector X′ i User trajectory T A The vector corresponding to the m-th check-in position, tanh() is the activation function, W l W g Both and w are training parameters. The final embedded representation vector s′ j The calculation formula is: s′ j =σ(W T [s j ;h j ]+b T ), a m,i =w T (tanh([W l X′ j，m ;W g h i ])), In the formula l j For user trajectory T B The length of x′ j，m To enhance the check-in location vector X′ j User trajectory T B The vector corresponding to the m-th check-in position.

6. The cross-social network user identity alignment method based on user spatiotemporal data according to claim 1, characterized in that: in, The matching probability The calculation formula is: In the formula, W1 and W2 are both training parameter matrices, and b1 and b2 are both bias terms.

7. The cross-social network user identity alignment method based on user spatiotemporal data according to claim 1, characterized in that: in, The user identity alignment prediction model is trained using a training dataset constructed from existing social network data and a loss function, Loss. The training dataset includes M pairs of samples, each pair consisting of a sample and its corresponding ground truth label. The formula for calculating the loss function Loss is as follows: In the formula The matching probability y is obtained by inputting the user identity alignment prediction model into the i-th sample. i Let be the true label corresponding to the i-th sample.

Images