System and method for identifying cross-social network user identities based on network structure

Abstract

The application provides a cross-social network user identity recognition system and application method based on a network structure, which comprises the following steps: algorithm framework design, a random walk method is used to capture the structural correlation between nodes and aligned nodes; a user identity alignment stage, an optimal transmission theory-based user identity alignment optimization method is used to realize accurate alignment; a training process, a cross-network fusion layer is designed to combine the network shared features of the entire aligned nodes and the specific features of each node, and a graph attention network is used to aggregate neighbor node information to enhance the discrimination ability between nodes. The application can effectively improve the accuracy and performance of user identity alignment, and can help to provide cross-platform recommendation and cross-platform user portrait.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of online social networks, and in particular to a cross-social network user identity recognition system and application method based on network structure. BACKGROUND

[0002] With the popularity of online social platforms, various types of social platforms have emerged, and have gradually become an indispensable part of people's daily life. Due to the differences in function positioning, service scenarios and user interaction methods of different platforms, users usually have accounts on multiple platforms or services. For example, a person may have different accounts on social platforms (such as Facebook, Twitter, Weibo), e-commerce platforms (such as Taobao, Amazon), etc. These accounts exist independently on each platform, and lack a direct identity association mechanism. In this context, the demand for cross-social network user identity alignment has grown significantly, which aims to identify and match different platform accounts belonging to the same real user, in order to integrate and analyze cross-platform user information. Identity alignment not only helps to break through user behavior data and improve the quality of personalized services, but also has important value in building user portraits, optimizing recommendation systems, fraud detection, social network evolution analysis and many other scenarios. Therefore, researching effective cross-platform user identity alignment methods is of great significance for promoting the fusion analysis of multi-source heterogeneous social data.

[0003] In the past, the cross-social network user identity alignment problem was modeled by using available information such as user attributes or network structure. However, on different platforms, even for the same user, the attribute information often has inconsistencies and incompleteness, and may even contain false information, making it still a great challenge to obtain enough semantic information to accurately align user identities. In comparison, social network structure is a more robust feature, and the most direct approach for a method based on social network structure is to measure the structural similarity between aligned users, or to project different networks into a unified low-dimensional latent space, and by minimizing the distance between aligned user representations, to infer other potential aligned users. However, known aligned users may not have similar topological features in structure, on the contrary, the topological representation of a node in its network is often more similar, which will lead to misalignment in the alignment process. In recent years, optimal transport methods have been widely used to solve this problem and have achieved significant performance improvement. They mainly construct node feature consistency matrix and node structure consistency matrix as transport cost matrix respectively, and calculate Gromov-Wasserstein distance as the transport cost to be optimized based on this. Although these methods perform well in performance, multiple distance calculations and optimization processes often come with high computational cost. Therefore, a technical method is needed to overcome the above problems. SUMMARY

[0004] In view of the above problems,

[0005] One of the purposes of the present application is to provide a network structure-based cross-social network user identity alignment and optimization method.

[0006] The second purpose of the present application is to provide a network structure-based cross-social network user identity recognition system, which can effectively improve the accuracy and performance of user identity alignment by alleviating the influence of similar nodes on the prediction results and the prediction efficiency problem.

[0007] The third purpose of the present application is to provide an application method of the above-mentioned network structure-based cross-social network user identity recognition system, which can provide help for cross-platform recommendation and cross-platform user portrait.

[0008] A network structure-based cross-social network user identity alignment and optimization method, including algorithm framework design, training process, and user identity alignment stage:

[0009] I. Algorithm framework design, including:

[0010] Construct the graph structure of the source network and the target network, use the random walk method to capture the structural correlation between the nodes and the alignment nodes, and construct the initialization embedding representation of the nodes;

[0011] Introduce a decoupling mechanism to decouple the node representation into network-specific representation and network-shared representation, perform cross-network feature fusion based on the shared representation to enhance the semantic consistency of the alignment nodes, use graph attention network to weight and aggregate the neighborhood features of the nodes, and design a multi-loss optimization strategy for training optimization, including designing a reconstruction loss to ensure the expression ability of the node shared representation, and designing a cross-network contrast loss and a network-internal contrast loss to reduce the representation distance of the alignment nodes in the embedding space, while distinguishing similar nodes within the network to prevent misalignment.

[0012] II. In the user identity alignment stage:

[0013] The user identity alignment optimization method based on optimal transport theory is used to realize accurate alignment. The present application realizes this method on three real social network datasets, which proves that this method can significantly improve the accuracy and robustness of predicting user identity alignment, and provides better cross-platform recommendation services for users.

[0014] The above process, detailed steps, include the following:

[0015] S1, divide the data set into training set, test set and validation set, and construct the graph structure of the source network and the target network based on the online social network data set;

[0016] S2, initialize node embedding representation;

[0017] S3, decouple node representation;

[0018] S4, fuse and align node features;

[0019] S5, aggregate node neighborhood features and extract node embedding representation;

[0020] S6, design a multi-loss optimization strategy, and use the training set in S1 to train and optimize;

[0021] S7, based on the optimal transmission optimization username identity matching result.

[0022] A cross-social network user identity recognition system based on network structure, comprising:

[0023] The first module divides the open source data set into training set, test set and validation set, and constructs the graph structure of the source network and the target network of the online social network data set;

[0024] The second module is used for initializing node embedding representation;

[0025] The third module is used for decoupling node representation;

[0026] The fourth module is used for fusing and aligning node features;

[0027] The fifth module is used for aggregating node neighborhood features and extracting node embedding representation;

[0028] The sixth module is based on a multi-loss optimization strategy, and uses the training set in the first module to train and optimize;

[0029] The seventh module is based on the optimal transmission optimization username identity matching result.

[0030] An application method of a cross-social network user identity recognition system based on network structure, comprising:

[0031] Step 1: Select two social platforms from a plurality of social platforms as social platforms that need to collect data, collect the key information of users in the platforms respectively through a crawler system, including the relationship of the user, the username information, the text information generated by the user and the location information; store the information collected by the crawler system in the database, and perform data cleaning operation, remove redundant data, delete data with missing values, delete invalid abnormal data or sensitive information data;

[0032] Step 2: input the user in one platform into the cross-social network user identity recognition system based on network structure, and get the user aligned in another platform;

[0033] Step 3: The user data in different platforms is fused, the interest topics are extracted from the text published by the user by using a topic model, the behavior trajectory is described through time series analysis, and the clustering algorithm is used for clustering analysis of the interest topics and the behavior mode, so that a complete and accurate user feature portrait is constructed, including the interest preference and the activity area information;

[0034] Step 4: The fused user data is input into the cross-platform recommendation system based on collaborative filtering, the hybrid collaborative filtering method is used, the multi-dimensional similarity relationship between the user-user, the content-content and the user-content is constructed, the latent feature model based on matrix decomposition is combined, the implicit user interest feature and the content feature are further extracted, the effective migration of the cross-platform user preference is realized. The mechanism can establish the interest mapping relationship between the source platform and the target platform, generate more accurate, personalized and diversified recommendation results, and significantly improve the accuracy of the recommendation system and the user experience.

[0035] Compared with the prior art, the present application has the following beneficial effects:

[0036] The present application considers the problem of difficult differentiation of similar nodes in the cross-social network user identity alignment method based on network structure, and proposes a method, system and application which simultaneously considers alignment accuracy and prediction efficiency. The present application proposes a representation decoupling mechanism, maps the node representation to the network-specific space and the network-shared space, introduces a reconstruction loss to constrain the semantic integrity, and designs a cross-network fusion layer to fuse the network-shared features of the aligned nodes. Subsequently, the network-shared features are combined with the respective specific features, and the graph attention network is used to aggregate the neighbor node information to enhance the discrimination ability between nodes. In the node alignment stage, the present application constructs an alignment-guided node consistency matrix as the overall transmission cost matrix based on the node embedding representation and the local structure information of the aligned user, solves the possibility probability matrix of user alignment by minimizing the overall transmission cost, and realizes efficient and accurate user identity alignment. The above method improves the performance and robustness of the present application in user identity alignment, reduces the risk of misalignment, and takes into account the computational complexity. The system application of the present application can establish the interest mapping relationship between the source platform and the target platform, generate more accurate, personalized and diversified recommendation results, and significantly improve the accuracy of the recommendation system and the user experience. BRIEF DESCRIPTION OF DRAWINGS

[0037] Figure 1 The method principle of the present application is shown, wherein (a) is a framework diagram of the algorithm, (b) is the implementation details of the fused node feature, and (c) is the alignment optimization process based on OT

[0038] Figure 2 The inference time of the user identity alignment method on three data sets is shown

[0039] Figure 3 The application shows a network structure-based cross-social network user identity recognition system block diagram

[0040] Figure 4 The application shows a cross-platform recommendation system application diagram DETAILED DESCRIPTION

[0041] The application will be further described in detail below with reference to the accompanying drawings, so that those skilled in the art can implement the application according to the description and drawings.

[0042] First part algorithm design

[0043] As shown in the figure, a network structure-based cross-social network user identity alignment and optimization method comprises: Figure 1

[0044] S1, divide the data set into training set, test set, validation set, and construct the graph structure of the source network and the target network from the online social network data set;

[0045] S2, initialize node embedding representation;

[0046] S3, decouple node representation;

[0047] S4, fuse and align node features;

[0048] S5, aggregate node neighborhood features and extract node embedding representation;

[0049] S6, design a multiple loss optimization strategy, and use the training set of S1 to train and optimize;

[0050] S7, based on optimal transport, optimize the username identity matching result.

[0051] S1, divide the data set into training set, test set, validation set, and construct the graph structure of the source network and the target network from the online social network data set:

[0052] S1.1 three pairs of real online social network data sets, including Facebook-Twitter data set, Facebook-Twitter data set, and Douban-Microblog data set. For all aligned node pairs in the data set, randomly select 20%, 70%, and 10% of the aligned user pairs as training samples, test samples, and validation samples, respectively, to construct the training set sample set , test set sample set , and validation set sample set .

[0053] The selected three pairs of online social network data sets are:

[0054] ​The foursquare-Twitter dataset (extracted from “Zhang J, Yu PS. Integrated anchor and social link predictions across social networks [C] / / Proceedings of the 24th International Joint Conference on Artificial Intelligence. 2015: 2125-2131.”), where the foursquare platform contains 5312 users and 76920 user relationships, the Twitter platform includes 5120 users and 164920 user relationships, and the number of aligned users is 3138.

[0055] The Facebook-Twitter dataset (extracted from “Xuezhi Cao and Yong Yu. ASNets: A benchmark dataset of aligned social networks for cross-platform user modeling. [C] / / In Proceedings of the 25th ACM International on Conference on Information and Knowledge Management, 2016: 1881-1884.”), where the Facebook platform contains 2458 users and 40298 user relationships, the Twitter platform includes 2458 users and 95034 user relationships, and the number of aligned users is 2458.

[0056] The Douban-Sina dataset (extracted from “Xuezhi Cao and Yong Yu. ASNets: A benchmark dataset of aligned social networks for cross-platform user modeling. [C] / / In Proceedings of the 25th ACM International on Conference on Information and Knowledge Management, 2016: 1881-1884.”), where the Facebook platform contains 3154 users and 301074 user relationships, the Twitter platform includes 3154 users and 241736 user relationships, and the number of aligned users is 3154.

[0057] S1.2 In the cross-social online social network dataset, based on user identity alignment, the connection relationship between users of each network, the source network is constructed and the target network , wherein: the source network includes a set of user nodes from the source network and an adjacency relationship matrix between nodes , the target network includes a set of user nodes from the target network and an adjacency relationship matrix between nodes , is a real set, and the above adjacency relationship matrix represents the connection relationship between nodes.

[0058] S1.3 Based on the positive sample pairs in the training set, that is, the known aligned node pairs, an aligned node set is constructed , and a user alignment matrix is constructed accordingly , if any node i in the source network and node j in the target network are aligned users, that is, , then , otherwise . For each pair of aligned users , 10 non-aligned users are randomly selected from the neighborhood of user as negative samples for user u.

[0059] S2, initialize node embedding representation:

[0060] S2.1 Based on the source network and the target network constructed in S1, two different restart random walk (RWR) strategies RWR1 and RWR2 are used respectively to construct the relative position between each node and the known aligned anchor user (u, ).

[0061] S2.2 Calculate the RWR score based on the iteration process of RWR , and aggregate the RWR score as the initialized node representation .

[0062] The calculation formula of the RWR score of each node is:

[0063] (1)

[0064] In formula (1)

[0065] , indicates the source network or the target network .

[0066] is the restart probability;

[0067] is a one-hot vector with the u-th position as 1 and other positions as 0;

[0068] is the row-normalized adjacency matrix, where is the adjacency matrix of the network, and D is the diagonal matrix, denotes the transpose operation.

[0069] S3, decoupled node representation

[0070] S3.1 initialization of S2 node representation are input into multi-layer perceptron (MLP) respectively, which are mapped to network-shared feature space and network-specific feature space, and are decoupled into shared feature and network-specific feature , the calculation formulas are respectively:

[0071] (2)

[0072] (3)

[0073] In formula (2), (3)

[0074] denotes the source network or the target network ;

[0075] and are two different multi-layer perceptrons, which are composed of two linear layers and ReLU() functions.

[0076] S3.2 based on decoupled network-shared feature , the adjacency matrix of the source network and the target network is reconstructed respectively, and the calculation formula is:

[0077] (4)

[0078] S3.3 introduce the reconstruction loss function in S1 source network and target network respectively to further constrain the semantic integrity of the decoupled node feature, and the calculation formula of the reconstruction loss of the whole method is:

[0079] = + (5)

[0080] In formula (5) and are reconstruction losses of source network and target network , and the calculation formula of the reconstruction loss of each network is as follows:

[0081] (6)

[0082] In formula (6)

[0083] denotes an adjacency matrix of the reconstruction network,

[0084] denotes an adjacency matrix of the original network,

[0085] , denotes the source network or the target network .

[0086] S4, fusion alignment node features:

[0087] S4.1 based on known alignment node set and shared features The application adopts a cross-network representation fusion mechanism, improves the consistency of node representation by fusing the alignment node features from different networks, and obtains the fused node representation , wherein the calculation formula of the fused node representation of each pair of alignment nodes is as follows:

[0088] (7)

[0089] In formula (7)

[0090] and denote trainable parameter matrices,

[0091] denotes a Sigmoid function,

[0092] and denote bias terms,

[0093] denotes an element-level multiplication operation,

[0094] Node u and node v represent a known alignment node pair, denotes the source network The shared eigenvector of the middle node u, The target network The middle node The shared eigenvector of the middle node.

[0095] S4.2 Fused node representation Spliced with its network-specific features After splicing, the fused node representation is obtained The calculation formula is:

[0096] (8)

[0097] In formula (8) Indicates the splicing operation.

[0098] S5, aggregate node neighborhood features and extract node embedding representation, wherein:

[0099] S5.1 Based on the fused node features across networks in S4 The original graph is input into the graph attention network (GAT), and the neighborhood information of the node is aggregated to the node i according to the importance of different neighbors If any node i is taken as the center, the node representation aggregated with neighborhood information is obtained , Indicates the number of nodes in the network, and the calculation formula of each node representation is:

[0100] (9)

[0101] In formula (9)

[0102] Indicates the Elu activation function;

[0103] Indicates the neighbor node set of node i;

[0104] Indicates the trainable weight parameter when the node is aggregated in the graph attention network;

[0105] Indicates the attention coefficient between node i and its neighbor node j, which is calculated according to the correlation between node i and its neighbor node, and the calculation formula is: (10)

[0106] In formula (10)

[0107] Indicates the exponential function, Indicates the trainable vector, Indicates the trainable weight parameter when calculating the attention coefficient, Represents a nonlinear function. This indicates a splicing operation. This represents the node representation vector of node i after cross-network fusion.

[0108] S5.2 will reconstruct the source network. and reconstruct the target network The input is fed into a graph attention network (GAT), and step S5.1 is repeated. By calculating the attention coefficients, the neighbor information of the center node is aggregated to obtain the node representation with aggregated neighborhood information. .

[0109] S6, Based on the known set of alignment nodes in S1 Multiple loss functions are introduced to optimize training parameters;

[0110] The multiple losses include cross-network comparison loss. Contrast loss within the network Reconstruction loss Therefore, the designed multiple loss function The calculation formula is:

[0111] (11)

[0112] In formula (11) and It is a hyperparameter that balances the reconstruction loss and the intra-network contrastive loss. A cross-social network user identity matching method based on network structure continuously minimizes the loss function and backpropagates to update the model's parameters until the model converges or reaches the maximum number of iterations, at which point training stops.

[0113] Based on the known set of alignment nodes in S1 Cross-network contrastive loss is introduced from the perspectives of the original image and the reconstructed image, respectively. This reduces the distance between aligned nodes in the embedding space and increases the distance between unaligned nodes during training, ultimately improving the cross-network contrastive loss. = + The original image cross-network contrast loss The calculation formula is:

[0114] (12)

[0115] In equation (12) Let represent the similarity function for aligned nodes, where This represents the temperature coefficient parameter. This represents the set of aligned nodes used for training. Let i represent the set of negative samples for node i. Indicates a negative sample. and The node representation of the aggregated neighborhood information of node i and node j.

[0116] Similarly, the node representation of the aggregated neighborhood information of the reconstructed graph is used. and The cross-network contrastive loss is obtained by using it as input to the cross-network contrastive loss of the reconstructed graph. .

[0117] Based on source network and target network The representations of nodes in the source network and their corresponding representations in the reconstructed network are compared in the source network. and target network The embedding representation of the same node in the network, distinguishing the features of the same node from those of other nodes to prevent misalignment, and using contrastive loss within the network. The contrast loss within each network The calculation formula is:

[0118] (13)

[0119] In formula (13) Let i represent the set of negative samples for node i. This indicates a negative sample.

[0120] Based on the construction of the reconstructed adjacency matrix and the calculation method of the reconstruction loss described in S3.2 and S3.3, the reconstruction loss is obtained. .

[0121] S7. Optimize username and identity matching results based on optimal transmission:

[0122] S7.1 After the first part of training is completed, this invention is based on the learned node features. The transmission cost matrix is ​​constructed by calculating the embedding similarity and structural similarity of aligned nodes between networks. Its calculation formula

[0123] for:

[0124] (14)

[0125] In formula (14) Represents Euclidean distance. Represents an exponential function. This indicates the structural similarity of aligned nodes. This represents a known user alignment matrix. and Representing the source network respectively and target network The adjacency matrix.

[0126] S7.2 The optimization objective function is to minimize the Wasserstein Distance (WD). Measure the source network and target network The nodes represent distribution differences, and the probability matrix T of user alignment is solved using the Sinkhorn iterative algorithm, based on which the source network is found. Nodes in the target network Alignment nodes on, where the objective function to be optimized The calculation formula is:

[0127] (15)

[0128] In formula (15) Let be the probability of alignment between node i and node j. Represents the transmission matrix. and Representing the source network respectively and target network The set of nodes.

[0129] S7.3 The method is evaluated using the Top-K evaluation metrics Hits@K and the mean reciprocal ranking (MRR) based on the probability matrix T of user alignment.

[0130] The calculation formulas for the evaluation metrics Hits@K and the mean reciprocal ranking (MRR) are as follows:

[0131] (16)

[0132] (17)

[0133] In equations (16) and (17) This represents the test set. This indicates the number of samples in the test set. This indicates the correct order of users corresponding to user node i. {} indicates an indicator function.

[0134] Verification experiments were conducted to further demonstrate the technical effectiveness of the algorithm solution proposed in this application.

[0135] Select the test set sample set in S1 This method is used to verify its performance on real online social networks.

[0136] To verify the performance of the method of the present invention, the verification test of the present invention includes: (1) performance comparison of user identity alignment methods; (2) impact of the strategy proposed in the present invention on the prediction results; and (3) comparison of prediction time of user identity alignment methods.

[0137] The specific experimental setup is as follows:

[0138] All experiments used the Top-K evaluation metrics Hits@K and Mean Reciprocal Ranking (MRR). The mini-batch size was set to 256, the learning rate to 0.001, and K was set to 1, 10, and 30. The Adam optimizer was used as the optimizer for the entire experiment.

[0139] (1) Performance comparison of user identity alignment methods

[0140] Ten different network-based user identity alignment methods, including Deep Link, DANA, BRIGHT, NeXtAlign, CCNE, CINA, ASSISTANT, PARROT, HOT, and JOENA, were used as comparison methods, and the models were trained on three datasets respectively.

[0141] The performance comparison of the network structure-based user identity alignment method on three datasets is shown in Table 1.

[0142] Table 1:

[0143]

[0144] The experimental results show that the prediction results on all three datasets significantly outperform other comparative methods across all four evaluation metrics. Specifically, in terms of the MRR metric, this invention achieves improvements of 11.8%, 8.91%, and 2.61% on the three datasets, respectively. Particularly on the relatively sparse Four-Way Network-Twitter dataset, this demonstrates that the invention can more effectively capture the potential similarities between aligned nodes, thereby identifying more potential aligned users. This further validates the excellent performance of this invention in user identity alignment tasks, especially its more stable performance when handling sparse datasets.

[0145] (2) The impact of the strategy of this invention on the prediction results

[0146] w / o-RDL: Removes decoupled node features and directly uses the node representation generated by Restart Random Walk (RWR) as the input to the cross-network fusion layer;

[0147] w / o-DCRF: Removes the fused and aligned node features and directly inputs the decoupled node features into the graph attention network;

[0148] w / o-Recon: Remove the reconstruction loss term;

[0149] w / o-Intra-CL: Removes the in-network contrast loss term.

[0150] The impact of different strategies on prediction results on three datasets is shown in Table 2.

[0151] Table 2:

[0152]

[0153] The experimental results show that the present invention outperforms the aforementioned variant methods in terms of performance. The present invention outperforms all variants on all three datasets, verifying the effectiveness of each component of the system. In the characterization decoupling module, both removing RDL (decoupling node features) and DCRF (fusion alignment node features) lead to a performance decrease, indicating that both play a key role in identifying aligned users.

[0154] Specifically, removing DCRF results in a more significant performance drop on MRR and Hits@1 metrics, indicating that this module effectively improves the model's ability to achieve precise matching. However, on the Hits@30 metric, w / o-DCRF outperforms w / o-RDL, suggesting that RDL is more helpful in capturing potential similar features, thereby improving alignment. Regarding the loss term, removing the reconstruction loss results in a slight performance drop, indicating that preserving the structural information between nodes has a certain auxiliary effect on user alignment. On the other hand, removing the intra-network contrast loss leads to a more significant performance decline, highlighting the importance of improving node representation discrimination ability and modeling local structural patterns for the alignment task.

[0155] (3) Comparison of prediction time of user identity alignment methods

[0156] The inference phase runtime of the method of this invention is compared with that of a network-based user identity alignment method, and the model is trained on three datasets respectively.

[0157] Experimental results are as follows Figure 2 As shown.

[0158] The experimental results show that the method of this invention further reduces inference time on all three datasets, achieving lower time overhead while improving alignment performance.

[0159] For Table 1, Table 2, Figure 2 Explanation of the English information appearing in the text:

[0160] FourNet-Twitter Dataset: FourNet is a location-based mobile social networking platform, and Twitter is a popular microblogging social service platform. This dataset contains 3138 aligned nodes.

[0161] Facebook-Twitter dataset: This dataset contains 2458 aligned nodes. Facebook is a popular and highly influential social media service platform.

[0162] Douban-Weibo: Douban is a social platform based on users' interests in movies and books, while Weibo is a popular and leading social media platform in China that allows people to create, share, and discover blogs. This dataset contains 3154 aligned nodes.

[0163] Hits@K represents the percentage of true aligned nodes that appear in the predictions of the first K aligned nodes out of all tested aligned nodes. In the experimental validation, K was set to 1, 10, and 30, respectively.

[0164] The mean reciprocal rank (MRR) is a measure of the average reciprocal rank of the true aligned nodes.

[0165] The comparative methods used in the experiment included:

[0166] A deep learning method for user identity linking (DeepLink), a domain adversarial network alignment method (DANA), a linking algorithm for network alignment (BRIGHT), a method for balancing consistency and dissimilarity in network alignment (NeXtAlign), a collaborative cross-network embedding framework method for network alignment (CCNE), a curvature-based hypergraph integration network alignment method (CINA), an efficient network alignment method utilizing trusted node attributes (ASSISTANT), a location information-aware regularized optimal transmission method for network alignment (PARROT), a hierarchical multilateral optimal transmission method for network alignment (HOT), a joint optimal transmission and embedding method for network alignment (JOENA), and a cross-network user identity alignment and optimization method based on network structure (DCLO).

[0167] w / o-RDL: Removes decoupled node features and directly uses the node representation generated by Restart Random Walk (RWR) as the input to the cross-network fusion layer;

[0168] w / o-DCRF: Removes the fused and aligned node features and directly inputs the decoupled node features into the graph attention network;

[0169] w / o-Recon: Remove the reconstruction loss term;

[0170] w / o-Intra-CL: Removes the in-network contrast loss term.

[0171] Part Two

[0172] Based on the algorithm of this invention, a cross-social network user identification system based on network structure and its application method are further presented.

[0173] like Figure 3 As shown, a cross-social network user identification system based on network structure includes the following:

[0174] Module 1: Divide the open-source dataset into training set, test set, and validation set, and construct the graph structure of the source network and target network from the online social network dataset;

[0175] Module 2 is used to initialize the node embedding representation;

[0176] Module 3 is used to decouple node representations;

[0177] Module 4, used for fusing and aligning node features;

[0178] Module 5 is used to aggregate node neighborhood features and extract node embedding representations;

[0179] Module 6. Based on the multiple loss optimization strategy, use the training set described in Module 1 for training optimization;

[0180] Module 7. Optimize username and identity matching results based on optimal transmission.

[0181] The first module 1 includes three sub-modules: first module sub-module one, first module sub-module two, and first module sub-module three, wherein:

[0182] In the first module, submodule one, for all aligned node pairs in the dataset, randomly selects 20%, 70%, and 10% of the aligned user pairs as training samples, test samples, and validation samples, respectively, to construct the training sample set. Test set sample set and validation set sample set ;

[0183] In the first module, submodule two, based on the cross-social online social network dataset, the source network is constructed by aligning user identities and determining the connection relationships between users in each network. and target network Wherein: the source network Includes the set of user nodes from the source network Adjacency matrix between nodes The target network Includes a set of user nodes from the target network. Adjacency matrix between nodes , Given a set of real numbers, the adjacency matrix above represents the connection relationships between nodes;

[0184] Sub-module 3 of the first module constructs an alignment node set based on the positive sample pairs in the training set, i.e., the known alignment node pairs. And based on this, construct a user alignment matrix. If any node i in the source network and node j in the target network are aligned users, i.e. ,but ,otherwise For each pair of aligned users User u in the target network is randomly selected from users. Ten unaligned users are selected from the neighborhood of the selected user as its negative samples.

[0185] The second module 2 is used to initialize the node embedding representation, including second module sub-module one and second module sub-module two, wherein:

[0186] Sub-module 1 of the second module is based on the source network built in module 1. and target network Two different restart random walk (RWR) strategies, RWR1 and RWR2, are used to construct a connection between each node and a known alignment anchor user (u, The relative positions between them;

[0187] Sub-module 2 of module 2: Calculating RWR scores based on the iterative process of RWR And by aggregating the RWR scores as the initial node representation. ;

[0188] The formula for calculating the RWR score for each node is as follows:

[0189] (1)

[0190] In formula (1)

[0191] , indicating the source network or target network Two scenarios;

[0192] , is the restart probability;

[0193] This represents a one-hot vector starting from the u-th position, where the u-th position is 1 and all other positions are 0.

[0194] is a row-regularized adjacency matrix, where Let D represent the adjacency matrix of the network, and let D represent the angle matrix. This indicates the transpose operation.

[0195] The third module 3 is used to decouple node representations, and includes three sub-modules: sub-module 1, sub-module 2, and sub-module 3.

[0196] The third module's sub-module one represents the nodes initialized in the second module 2. The inputs are fed into a multilayer perceptron (MLP) and mapped to the network shared feature space and the network specific feature space, thus decoupling them into shared features. and specific characteristics of the network The calculation formulas are as follows:

[0197] (2)

[0198] (3)

[0199] In equations (2) and (3)

[0200] , indicating the source network or target network Two scenarios;

[0201] and These are two different multilayer perceptrons, both consisting of two linear layers and the ReLU() function;

[0202] Sub-module 2 of the third module: Network sharing features based on decoupling Reconstruct the source network respectively and target network adjacency matrix The calculation formula is:

[0203] (4)

[0204] The third module, sub-module three, respectively contains the source network in the first module 1. and target network A reconstruction loss function is introduced to constrain the semantic integrity of the features of decoupled nodes, thus increasing the overall reconstruction loss of the method. The calculation formula is:

[0205] = + (5)

[0206] In formula (5) and Source network and target network The reconstruction loss is calculated using the following formula for each network:

[0207] (6)

[0208] In formula (6)

[0209] This represents the adjacency matrix of the reconstructed network.

[0210] This represents the adjacency matrix of the original network.

[0211] , indicating the source network or target network Two scenarios.

[0212] The fourth module 4 is used to fuse alignment node features, and includes fourth module sub-module one and fourth module sub-module two, wherein:

[0213] Sub-module 1 of the fourth module is based on the known set of alignment nodes. and shared features A cross-network representation fusion mechanism is adopted to improve the consistency of node representations by fusing aligned node features from different networks, resulting in a fused node representation. Each pair of aligned nodes Fusion node representation The calculation formula is:

[0214] (7)

[0215] In equation (7)

[0216] and Represents the trainable parameter matrix.

[0217] This represents the Sigmoid function.

[0218] and Indicates the bias term.

[0219] This represents element-wise multiplication.

[0220] Nodes u and v represent known aligned node pairs. Indicates source network The shared feature vector of node u in the middle. Represents the target network Middle node Shared feature vectors;

[0221] Node representation of the fusion of submodule 2 in module 4 Its specific network characteristics After splicing, the merged node representation is obtained. The calculation formula is:

[0222] (8) In equation (8) This indicates a splicing operation.

[0223] The fifth module 5 is used to aggregate node neighborhood features and extract node embedding representations, including fifth module sub-module one and fifth module sub-module two, wherein:

[0224] Sub-module 1 of module 5 is based on the node features after cross-network fusion in module 4. The original graph is input into a Graph Attention Network (GAT), where the neighborhood information of nodes is processed through different attention coefficients. If we take any node i as the center and aggregate its different neighbors according to their importance, we obtain a node representation that aggregates neighborhood information. , This represents the number of nodes in the network, where the formula for calculating the representation of each node is:

[0225] (9)

[0226] In equation (9)

[0227] Represents the Elu activation function;

[0228] Represents the set of neighboring nodes of node i;

[0229] This represents the trainable weight parameters used when nodes are aggregated in a graph attention network.

[0230] Let represent the attention coefficient between node i and its neighbor node j. It is calculated based on the correlation between node i and its neighbors, and its formula is as follows: (10)

[0231] In equation (10),

[0232] Represents an exponential function. This represents a trainable vector. This represents the trainable weight parameters used in calculating the attention coefficient. Represents a nonlinear function. This indicates a splicing operation. Represents the node representation vector of node i after cross-network fusion;

[0233] Submodule 2 of module 5 will reconstruct the source network. and reconstruct the target network The input is fed into a Graph Attention Network (GAT), and the fifth module submodule one is repeatedly called and run. By calculating the attention coefficients and aggregating the neighbor information of the center node, the node representation with aggregated neighborhood information is obtained. .

[0234] The sixth module 6 is based on the set of alignment nodes known from the first module 1. Multiple loss functions are introduced to optimize training parameters;

[0235] The multiple losses include cross-network comparison loss. Contrast loss within the network Reconstruction loss Multiple loss functions The calculation formula is:

[0236] (11)

[0237] In formula (11) and It is a hyperparameter that balances reconstruction loss and intra-network contrastive loss; a cross-social network user identity matching method based on network structure continuously minimizes the loss function and backpropagates to update the model's parameters until the model converges or reaches the maximum number of iterations, at which point training is terminated.

[0238] The sixth module 6:

[0239] Based on the known set of alignment nodes in Module 1 Cross-network contrastive loss is introduced from the perspectives of the original image and the reconstructed image, respectively. This reduces the distance between aligned nodes in the embedding space and increases the distance between unaligned nodes during training, ultimately improving the cross-network contrastive loss. = + The original image cross-network contrast loss The calculation formula is:

[0240] (12)

[0241] In equation (12) Let represent the similarity function for aligned nodes, where This represents the temperature coefficient parameter. This represents the set of aligned nodes used for training. Let i represent the set of negative samples for node i. Indicates a negative sample. and A node representation that represents the aggregated neighborhood information of node i and node j;

[0242] Similarly, the node representation of the aggregated neighborhood information of the reconstructed graph and The cross-network contrastive loss is obtained by using it as input to the cross-network contrastive loss of the reconstructed graph. .

[0243] The sixth module 6:

[0244] Based on source network and target network The representations of nodes in the source network and their corresponding representations in the reconstructed network are compared in the source network. and target network The embedding representation of the same node in the network, distinguishing the features of the same node from those of other nodes to prevent misalignment, and using contrastive loss within the network. The contrast loss within each network The calculation formula is:

[0245] (13)

[0246] In formula (13) Let i represent the set of negative samples for node i. This indicates a negative sample.

[0247] The seventh module 7 is used to optimize username and identity matching results based on optimal transmission, and includes sub-module 1, sub-module 2, and sub-module 3 of the seventh module, wherein:

[0248] After training is complete, the first submodule of the seventh module uses the learned node features... The transmission cost matrix is ​​constructed by calculating the embedding similarity and structural similarity of aligned nodes between networks. The calculation formula is as follows:

[0249] (14)

[0250] In formula (14) Represents Euclidean distance. Represents an exponential function. This indicates the structural similarity of aligned nodes. This represents a known user alignment matrix. and Representing the source network respectively and target network The adjacency matrix;

[0251] Submodule 2 of module 7 optimizes by minimizing the Wasserstein Distance (WD) as the objective function. Measure the source network and target network The nodes represent distribution differences, and the probability matrix T of user alignment is solved using the Sinkhorn iterative algorithm, based on which the source network is found. Nodes in the target network Alignment nodes on, where the objective function to be optimized The calculation formula is:

[0252] (15)

[0253] In formula (15) Let be the probability of alignment between node i and node j. Represents the transmission matrix. and Representing the source network respectively and target network The set of nodes;

[0254] Sub-module 3 of module 7 evaluates the method based on the probability matrix T of user alignment using the Top-K evaluation metrics Hits@K and the mean reciprocal ranking (MRR).

[0255] The calculation formulas for the evaluation metrics Hits@K and the mean reciprocal ranking (MRR) are as follows:

[0256] (16)

[0257] (17)

[0258] In equations (16) and (17) This represents the test set. This indicates the number of samples in the test set. This indicates the correct order of users corresponding to user node i. {} indicates an indicator function.

[0259] Application technology solution: Collect user behavior data and content from multiple platforms, and conduct cross-platform recommendation applications through user identification systems and recommendation systems.

[0260] like Figure 4 As shown, the application method of a cross-social network user identification system based on network structure is as follows:

[0261] Step 1: Select two social media platforms from among many as the platforms from which data needs to be collected. Use a web crawler (using the Selenium framework) to collect key user information from each platform, including user relationships, usernames, user-generated text information, and location information. Store the information collected by the web crawler in a database and perform data cleaning operations to remove redundant data, delete data with many missing values, and remove invalid, abnormal, and sensitive information.

[0262] Step 2: Input a user from one platform into a cross-social network user identification system (hereinafter referred to as the "user identification system") based on network structure to obtain the user on another platform that the user is aligned with.

[0263] Step 3: Merge user data from different aligned platforms, extract interest topics from user-posted texts using topic models, characterize user behavior trajectories through time series analysis, and perform clustering analysis on interest topics and behavioral patterns to construct a complete and accurate user profile, including information such as interest preferences and activity areas.

[0264] Step 4: Input the fused user data into a cross-platform recommendation system based on collaborative filtering. This system employs a hybrid collaborative filtering method, constructing multi-dimensional similarity relationships between users, content, and content, and combining this with a latent feature model based on matrix factorization to further extract implicit user interest features and content features, achieving effective transfer of user preferences across platforms. This mechanism can establish interest mapping relationships between the source and target platforms, generating more accurate, personalized, and diverse recommendation results, thereby significantly improving the accuracy of the recommendation system and the user experience.

Claims

1. A network structure based cross social network user identity recognition system, characterized in that, Comprise as follows: The first module (1), the open source data set is divided into training set, test set, validation set, and the online social network data set is constructed source network and target network graph structure; The second module (2) is used for initializing node embedding representation; The third module (3) is used for decoupling node representation; The fourth module (4) is used for fusing and aligning node features; The fifth module (5) is used for aggregating node neighborhood features and extracting node embedding representation; The sixth module (6) is based on multiple loss optimization strategy, and the training set in the first module (1) is used for training optimization; The seventh module (7) is based on optimal transport method to optimize the username identity matching result; The first module (1) comprises a first module submodule one, a first module submodule two and a first module submodule three, wherein: The first module submodule one randomly selects 20%, 70%, and 10% of the aligned user pairs in all aligned node pairs in the data set as training samples, test samples, and validation samples respectively to construct a training set sample set , a test set sample set , and a validation set sample set ​ The first module sub-module two constructs a source network and a target network based on user identity alignment and connection relationship between users of each network in a cross-social online social network data set , Wherein, the source network Comprises a user node set from the source network And an adjacency relationship matrix between nodes The target network Comprises a user node set from the target network And an adjacency relationship matrix between nodes , The adjacency relationship matrix is an adjacency relationship between nodes. The first module sub-module three is based on the positive sample pair in the training set, that is, the known alignment node pair, to construct an alignment node set , and to construct a user alignment matrix accordingly If any node i in the source network and node j in the target network are alignment users, that is, , then , otherwise ; for each pair of alignment users , 10 non-alignment users are randomly selected from the neighborhood of the user as negative samples of the user u. The third module (3) is used for decoupling node representation, comprising a third module submodule one, a third module submodule two and a third module submodule three, wherein: The third module sub-module one represents the node initialized by the second module (2) are input into a multi-layer perceptron (MLP) respectively, mapped to a network-shared feature space and a network-specific space, and decoupled into shared features and network-specific features , and the calculation formulas are as follows: （2） （3） In formula (2), (3) , indicating the source network or the target network both cases; and are two different multi-layer perceptrons, both consisting of two linear layers and ReLU() functions; Third module submodule two based on decoupling network sharing features , respectively, reconstruct the source network and target network adjacency matrix , the formula is: （4） The third module sub-module three respectively in the first module (1) source network and target network The reconstruction loss function is introduced to further constrain the semantic integrity of the decoupling node features, and the calculation formula of the reconstruction loss of the whole method is: ​ = + （5） In formula (5) and are reconstruction losses of source network and target network respectively, and the calculation formula of the reconstruction loss of each network is: （6） In formula (6) denotes the adjacency matrix of the reconstructed network, denotes the adjacency matrix of the original network, , indicates the source network or the target network in both cases; The fourth module (4) is used for fusing and aligning node features, comprising a fourth module submodule one and a fourth module submodule two, wherein: The fourth module sub-module one is based on a known alignment node set and shared features , adopts a cross-network representation fusion mechanism, improves the consistency of node representation by fusing alignment node features from different networks, and obtains a fused node representation wherein each pair of alignment nodes The calculation formula of the fused node representation is: （7） In formula (7) and denotes a trainable parameter matrix, denotes a sigmoid function, and denotes a bias term, represents an element-wise multiplication operation, node u and node v represent a known aligned pair of nodes, representing a source network shared feature vector of node u, representing a target network shared feature vector of node ​ Fourth module submodule two fused node representation With its network specific features After splicing, the fused node representation is obtained The calculation formula is: （8） In formula (8) represents a concatenation operation; The fifth module (5) is used for aggregating node neighborhood features and extracting node embedding representation, comprising a fifth module submodule one and a fifth module submodule two, wherein: The fifth module submodule one is based on the fourth module (4) across the network fusion node characteristics The original graph is input into a graph attention network (GAT), and the neighborhood information of a node is aggregated to the node i according to the importance of different neighbors , and the node representation aggregated with the neighborhood information is obtained , , wherein the calculation formula of each node representation is: （9） In formula (9) denotes the Elu activation function; N(i) denotes the set of neighbor nodes of node i; representing trainable weight parameters at the time of node aggregation in the graph attention network; denotes the attention coefficient between node i and its neighbor node j, which is calculated according to the relevance of node i and its neighbor node, and the calculation formula is: (10) In formula (10), denotes an exponential function, denotes a trainable vector, denotes a trainable weight parameter when computing the attention coefficient, denotes a non-linear function, denotes a concatenation operation, denotes the cross-network fused node representation vector for node i. The fifth module submodule two will reconstruct the source network and the target network into a graph attention network (GAT), repeatedly call and run the fifth module submodule one, aggregate the neighbor information of the center node by calculating the attention coefficient, and obtain the node representation of the aggregated neighborhood information .

2. The system according to claim 1, wherein the system is based on a network structure. The second module (2) is used for initializing node embedding representation, comprising a second module submodule one and a second module submodule two, wherein: A second module sub-module one based on the source network constructed by the first module (1) and the target network , respectively using two different restart random walk (RWR) strategy RWR1 and RWR2, construct each node and the relative position between the known alignment anchor point user (u, ) The second module submodule two calculates RWR scores based on an iterative process of RWR and by aggregating the RWR scores as initialized node representations ; The RWR score calculation formula of each node is: （1） In formula (1) , indicating the source network or the target network both cases; is the restart probability; denotes a one-hot vector with the u-th position as the starting point, i.e., the u-th position is 1, and the other positions are all 0; is the regularized adjacency matrix, where denotes the adjacency matrix of the network, D denotes the diagonal matrix, denotes the transpose operation.

3. The system of claim 1, wherein the system is based on a network structure. The sixth module (6) is based on the set of alignment nodes known by the first module (1) introducing multiple losses, optimizing training parameters; The multiple loss includes a cross-network contrast loss , a within-network contrast loss , a reconstruction loss , a multiple loss function The calculation formula is: （11） In formula (11) and is a hyperparameter that balances the reconstruction loss and the contrastive loss within the network; the loss function is constantly minimized, and the parameters of the model are updated by backpropagation until the model converges or the maximum number of iterations is reached, and the training is terminated.

4. The system according to claim 3, wherein the system is configured to: identify the user's social network identity based on the network structure. 5 The sixth module (6): Based on the first module (1) known alignment node set , respectively, in the original graph and the reconstructed graph angle introduces cross-network contrast loss, so that in the training process to reduce the distance of the alignment node in the embedding space, increase the distance of the non-alignment node, the final cross-network contrast loss = + , wherein the original graph cross-network contrast loss The calculation formula is: （12） In formula (12) represents a similarity function of aligned nodes, where represents a temperature coefficient parameter, represents a set of aligned nodes for training, represents a set of negative samples of node i, represents a negative sample, and represents a node representation of aggregated neighborhood information of node i and node j; Likewise, the node representation of the aggregated neighborhood information of the reconstructed graph and As the input of the cross-network contrastive loss, the cross-network contrastive loss of the reconstructed graph is obtained .

5. The system according to claim 3, wherein the system is configured to: identify the user's social network identity based on the network structure. 5 The sixth module (6): Based on the source network and the target network The representation of the nodes in the source network and the target network is compared with the representation of the nodes in the reconstructed network corresponding to the representation of the nodes in the source network and the target network , respectively, to distinguish the features of the same nodes from those of other nodes and prevent misalignment, and the intra-network comparison loss , wherein each intra-network comparison loss is calculated according to the following formula: （13） In formula (13) denotes the set of negative samples of node i, denotes a negative sample.

6. The system according to claim 3, wherein the system is configured to: identify the user's social network identity based on the network structure. The seventh module (7) is used for optimizing the username identity matching result based on the optimal transport method, comprising a seventh module submodule one, a seventh module submodule two and a seventh module submodule three, wherein: The seventh module submodule one calculates embedding similarity and structure similarity of aligned nodes between networks based on the learned node features after training is completed , and constructs a transmission cost matrix , and the calculation formula is: （14） In formula (14) denotes the Euclidean distance, denotes the exponential function, denotes the structural similarity of aligned nodes, denotes the known user alignment matrix, and denote the adjacency matrices of the source network and the target network respectively. The seventh module sub-module two takes the Wasserstein Distance (WD) as the optimization objective function by minimizing , which measures the difference of node representation distribution between the source network and the target network , and solves the possibility probability matrix T of user alignment by Sinkhorn iterative algorithm, and finds the aligned nodes of the nodes in the source network on the target network based on this, wherein the calculation formula of the optimization objective function is （15） In formula (15) is the probability of alignment between node i and node j, denotes the transmission matrix, and denote the node sets of the source network and the target network respectively. The seventh module submodule three is based on the possibility probability matrix T of user alignment, and adopts Top-K evaluation indexes Hits@K and average reciprocal rank (MRR) to evaluate the method; Wherein the calculation formula of evaluation index Hit@K and average reciprocal rank (MRR) is respectively: （16） （17） In formulas (16), (17) denotes the test set, denotes the number of samples in the test set, denotes the ranking of the correctly aligned user corresponding to the user node i, denotes the indicator function.

Images