A user identity association method for alleviating cross-platform heterogeneity

Abstract

The present application belongs to the field of data mining and association rules in information systems, and proposes a user identity association method to alleviate cross-platform heterogeneity. By designing a semantic perception representation alignment module, consistency between user-generated content on different platforms is established, effectively alleviating text attribute heterogeneity. And through the topological perception representation alignment module, the network densification mechanism is adopted to reduce the topological structure heterogeneity between networks. Both modules are trained based on the optimization strategy designed by the present application, which maximizes the similarity of anchor nodes while preserving the uniqueness of the neighborhood. The present application also introduces an adaptive representation integration module to dynamically calibrate the weight contribution of text attributes and topological structure information in user representation learning, thereby efficiently realizing cross-platform user identity association. The present application can be used as a basic module to provide accurate user identification and positioning support for downstream recommendation systems.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of data mining and association rules in information systems, and discloses a user identity linkage method for relieving cross-platform heterogeneity. BACKGROUND

[0002] The field of user identity linkage is rapidly developing, and its core goal is to match user accounts on different social platforms. Due to the wide value of this research in practical applications, the field has attracted much attention and research. Contemporary methods mainly model user identity linkage as an anchor node prediction problem across networks by analyzing the similarity of text attributes (such as user-generated content) or the relevance of social relationships, i.e., identifying the same user (anchor node pair) in different social networks. However, these methods face serious challenges in dealing with cross-platform heterogeneity and effectively integrating multiple information sources, especially when users exhibit specific attribute information and social interaction characteristics on different platforms.

[0003] Existing models face major challenges in the following three aspects:

[0004] 1. How to solve the problem of cross-network text attribute heterogeneity? The behavior patterns of the same user on different platforms often differ significantly. For example, a researcher may mainly publish papers related to natural language processing on natural language processing top conferences (such as EMNLP), while focusing on computer vision on computer vision conferences (such as ECCV), resulting in significant semantic differences. Current methods usually directly calculate similarity based on text attribute representation, making it difficult to effectively deal with this cross-network attribute heterogeneity.

[0005] 2. How to handle cross-network structural heterogeneity? The social connection patterns of users on different platforms often differ significantly. For example, a user's Google Scholar network reflects their academic collaboration relationships, while their connections on LinkedIn represent professional relationships. Existing methods usually learn social network topology attribute representations independently, making it difficult to capture cross-network structural differences and their impact on user identity linkage.

[0006] 3. How to effectively integrate text attribute information and topology structure information? User text attributes describe individual characteristics, while network structure reveals social interaction patterns, both of which have different importance in user identity linkage. Many current methods rely on simple concatenation of text attribute information and topology structure information, ignoring the different contribution levels of these heterogeneous information in user identity linkage.

[0007] In response to the above challenges, new solutions are urgently needed to better deal with cross-platform heterogeneity and make full use of multi-source information, promoting the development of the field of user identity linkage. SUMMARY

[0008] The present application is directed to the above-mentioned status and its deficiencies, and proposes a user identity association method to alleviate cross-platform heterogeneity. The method aims to provide a framework solution to effectively balance the importance of user text attributes and network topology features, while significantly reducing the impact of cross-platform heterogeneity on association accuracy.

[0009] The technical solution of the present application is a user identity association method to alleviate cross-platform heterogeneity. Through the semantic-aware representation alignment module and the topology-aware representation alignment module, combined with the adaptive integration mechanism of dynamic weighted text attribute representation and topology attribute representation, the cross-platform association of user identity is realized. Specifically, by designing the semantic-aware representation alignment module (Semantic-aware Representation Alignment) and the topology-aware representation alignment module (Topology-aware Representation Alignment), the feature representation capability is improved by using the consistency-enhanced learning mechanism (Consistency-enhanced Learning). Finally, the adaptive representation incorporation module (Adaptive Representation Incorporation) is introduced to generate the final representation, and anchor link prediction (Anchor Link Prediction) is performed, thereby efficiently realizing cross-platform user identity association. The present application can provide accurate user identification and positioning support for downstream recommendation systems.

[0010] The specific implementation steps are as follows:

[0011] S1 node representation learning: For the text attributes generated by users and the topology attributes of social networks, the present application uses the Google classic vector generation tool (Word2Vec) and the large-scale information network embedding method (LINE) to generate initial representations, respectively, representing the text attribute representation of the known anchor node and the social network topology attribute representation

[0012] S2 consistency-enhanced learning: Through the semantic-aware representation alignment module and the topology-aware representation alignment module, the extracted text attribute representation and topology attribute representation are aligned to enhance the consistency of the representation of the same user in different networks, and finally two enhanced representations are generated.

[0013] The detailed steps of S2 are as follows:

[0014] S2-1 inputs the extracted text attribute representation into the semantic-aware representation alignment module, according to the anchor nodes in the alignment problem, each anchor node pair (i.e., the representation of the same user in different networks) is considered a positive sample, while the text attribute representations of neighboring nodes are considered negative samples, thus enhancing the model's discriminative ability. This is achieved through the loss function L... attribute The optimization process ultimately generates an enhanced text attribute representation.

[0015] S2-2 supplements the neighborhood connections of anchor nodes identified as belonging to the same user in different networks through a bidirectional densification mechanism, reducing structural gaps while preserving the inherent characteristics of user relationships. Subsequently, the densified topological attribute representation is input into the topology-aware representation alignment module, using the same positive and negative sample selection strategy as S2-1, and through the loss function L... topology Optimization yields an enhanced topological attribute representation. The positive and negative sample selection strategies of the semantic awareness and topological awareness modules are consistent, but their parameters are independent of each other.

[0016] The S2-3 semantic-aware representation alignment module aims to distinguish different users while preserving the specific attribute characteristics of the same user; the topology-aware representation alignment module ensures that nodes maintain similar structural patterns in different networks.

[0017] The total loss function L for joint training of the two modules total =λ attribute ·L attribute +λ topology ·L topology L attribute and L topology These are the loss functions for the semantic-aware representation alignment module and the topology-aware representation alignment module, respectively, λ. attribute and λ topology These are the weighting coefficients of the two loss functions. The total loss function combines the optimization objectives of semantic and structural features.

[0018] S3 Adaptive Representation Integration: Although modules S2-1 and S2-2 achieve effective user identity association by preserving semantic and structural similarity, their independent processing of attribute and structural information fails to fully exploit their complementarity. This invention introduces an adaptive representation integration module to dynamically integrate textual attributes and topological structure information, thereby improving user identity association performance.

[0019] This module consists of two key parts: multiple expert networks and a router mechanism. Each expert network is a feedforward neural network that independently processes the input text attribute representation and topological attribute representation. The router dynamically calculates weights based on the input representations to determine the relevance of each expert network. The enhanced text content representations and topological content representations obtained in S2-1 and S2-2 are input into this adaptive representation integration module. After dynamic weighting by the expert networks and router, a weighted expert output node representation is output. The training loss of the adaptive representation integration model can be expressed as... in and These are anchor node pairs in different networks. The final node representation. When L ARI When the minimum value is reached, the distance between anchor node pairs in different networks is closest, thus yielding a further optimized anchor node correlation representation.

[0020] S4 Anchor Node Link Prediction: After obtaining a unified representation, this module identifies users across networks through distance calculation. Specifically, for anchor node pairs in two networks obtained from S3 (with final representation values ​​of [values ​​not specified] respectively), [the module] predicts the links based on distance. and The module will calculate their Euclidean distance as... The smaller the Euclidean distance, the greater the likelihood that the two nodes represent the same natural person.

[0021] Beneficial effects

[0022] This invention mitigates the heterogeneity of user-generated content (text attributes) and social networks (topological attributes) across different platforms and identifies user identities (i.e., anchor nodes). This invention provides a framework for linking user identities across networks, effectively addressing the challenges posed by attribute and structural heterogeneity in social networks.

[0023] This invention proposes a semantically aware representation alignment module, which successfully aligns the representations of user-generated content while preserving specific network features, effectively mitigating the impact of different content patterns across networks; it also introduces a topology-aware representation alignment module, which significantly enhances structural consistency while maintaining basic topology information through an innovative network densification mechanism.

[0024] Furthermore, this invention incorporates an adaptive representation integration architecture to dynamically integrate multi-perspective representations, providing a more flexible and efficient solution for representation fusion.

[0025] This invention significantly improves the accuracy of cross-network link prediction, that is, the precision of cross-network user identity association. This invention can serve as a foundational module and can be widely applied to the development of related services such as cross-network recommendation, user profiling, and malicious entity detection. Attached Figure Description

[0026] Figure 1 This is the training model structure for the present invention.

[0027] Figure 2 Comparison of experimental results. Detailed Implementation

[0028] The present invention will be further described below with reference to the accompanying drawings.

[0029] S1 node representation learning: The present application learns the feature representation of the user-generated text attributes and the topology attributes of the social network respectively. The present application uses the classic vector generation tool Word2Vec model of Google to represent the user-generated text information (such as published content, biography or related text attributes) for learning. As shown in Figure 1 , first, the text information is processed by word segmentation, which is decomposed into multiple words (word 1 to word N). Then, each word is mapped to a low-dimensional embedding vector by Word2Vec, and the average value of multiple word embedding vectors is aggregated, and finally the text representation is generated by the Softmax classifier. For the social network topology attributes of the anchor node, the large-scale information network embedding method (LINE) is used to generate the low-dimensional structure representation vector of the node according to the first-order similarity and the second-order similarity of the node.

[0030] S2 consistency enhanced learning: The present application inputs the text attribute representation extracted by S1 into the semantic perception representation alignment module, and inputs the topology attribute representation of the node into the topology perception representation alignment module. Through the semantic perception representation alignment module and the topology perception representation alignment module, the extracted text attribute representation and the topology attribute representation are aligned to enhance the consistency of the representation of the same user in different networks, and finally two enhanced representations are generated.

[0031] The detailed steps of S2 are as follows:

[0032] S2-1 inputs the text attribute representation into the semantic perception representation alignment module. Extract the text attribute representation from the anchor node pair (Train Anchors) of the source network and the target network. Each anchor node pair (i.e. the representation of the same user in different networks) is taken as a positive sample (anchor embeds1 and anchor embeds2 ), and the text attribute representation is selected from the neighbor nodes as a negative sample (neg embeds1 and neg embeds2 ), so as to improve the discriminant ability of the model to the consistency of the positive sample and the distinction of the negative sample. Through the loss function L attribute optimization of the module, the enhanced text attribute representation is finally generated.

[0033] S2-2 topology perception representation alignment module is similar to semantic perception module, but the topology structure of the graph is bidirectionally densified before processing. As shown in Figure 1 , given the anchor node A1 in the source network and the corresponding anchor node B1 in the target network, if A1 has a neighbor A2, and the corresponding anchor node B2 exists in the target network, but there is no connection with B1, a edge is established between B1 and B2. Subsequently, the topology attribute representation is taken as input, and the positive and negative samples are processed by batch division through GPU acceleration.

[0034] S2-3 uses a contrastive loss function to calculate the loss for positive and negative sample pairs, and accumulates the loss of each batch into the total loss (combined loss) to optimize the model's learning of positive sample consistency and negative sample discriminability. This is achieved through the loss function L... topology Optimization yields an enhanced topological attribute representation. The positive and negative sample selection strategies of the semantic awareness and topological awareness modules are consistent, but their parameters are independent of each other.

[0035] This invention achieves joint training by weighted summation of the loss functions of the semantic awareness and topology awareness modules. The total loss function is expressed as: L total =λ attribute ·L attribute +λ structure ·L topology .

[0036] S3 Adaptive Representation Integration: To further improve the effect of representation fusion, this invention introduces an adaptive representation integration module, which generates a joint representation through dynamic weight allocation and fusion.

[0037] This module consists of two key parts: multiple expert networks and router mechanisms;

[0038] Each expert network is a feedforward neural network that independently processes the input text attribute representation and topological attribute representation.

[0039] The router dynamically calculates weights based on the input representation to determine the relevance of each expert network;

[0040] The enhanced text content representation and topology content representation obtained from S2 are input into the adaptive representation integration module. After dynamic weighting, the expert network and router output a weighted expert output node representation.

[0041] The training loss of the adaptive representation integration model is expressed as: in and These are anchor node pairs in different networks. The final node representation, when L ARI When the minimum value is reached, the distance between anchor node pairs in different networks is closest, thus obtaining a further optimized anchor node correlation representation;

[0042] Each expert network consists of two fully connected layers (nn.Linear) and a nonlinear activation function (nn.ReLU) for text attribute representation or topological attribute representation. For router mechanisms, this invention uses fully connected layers for text attribute representation (x... a ) and topological attribute representation (x sEach expert's weight is generated separately and normalized using a softmax classifier to ensure the weights sum to 1. The Mixture of Experts Layer contains multiple expert networks (implemented via nn.ModuleList), which compute the outputs of all experts for each input text attribute representation and topological attribute representation, and then uses the weights generated by the router. a and weights s The weighted summation is performed to generate the final text attribute representation and topological attribute representation output. The weighted text attribute representation output and the topological attribute representation output are concatenated using torch.cat to form a joint representation. The adaptive layer (Mixture of Experts Sequential Layer) consists of multiple multi-expert fusion layers. Each layer includes input normalization (nn.LayerNorm), multi-expert fusion (Mixture of ExpertsLayer), output normalization (nn.LayerNorm), and residual connections (conbined). input +MoE output This improves training stability and generates the final joint representation.

[0043] S4 anchor node link prediction evaluates model performance using Top-K accuracy and Mean Reciprocal Rank (MRR). This invention measures the strength of association between nodes by calculating the Euclidean distance (or other similarity metric) between anchor node embeddings. All possible node pairs are ranked according to the calculation results; the smaller the distance, the stronger the predicted association. Top-K accuracy is calculated by checking whether the actual node pairs (positive samples) in the test set appear in the top K predicted node pairs, reflecting the model's accuracy in local predictions. The Mean Reciprocal Rank (MRR) is calculated for each positive sample in the test set, calculating its reciprocal rank (1 / Rank) in the prediction ranking, and averaging the reciprocal ranks of all positive samples to obtain the MRR. The MRR metric reflects the overall ranking performance of positive samples in the prediction results; a higher value indicates that the model can identify positive samples more quickly and accurately. By using both Top-K accuracy and MRR as evaluation criteria, the model's performance in local prediction accuracy (Top-K) and global ranking quality (MRR) can be comprehensively measured, ensuring that the model can not only effectively identify strongly correlated node pairs, but also reasonably rank the correlation strength of all node pairs.

[0044] Finally, the present application enhances the consistency of cross-network user identity association by integrating user text attribute representation and topology attribute representation, and improves the feature expression ability through adaptive fusion, thereby significantly improving the accuracy of user identity association prediction.

[0045] Finally, the cross-network link prediction score generated by the present application indicates the association of cross-network user identity. As shown in Figure 2 The experimental results of the present application (i.e., EFC-UIL in the figure) have the optimal performance compared with five baseline methods on three data sets. Among them, DBLP_1 and DBLP_2 are public data sets from the DBLP database (a database of computer science field collected journal and conference paper published papers), and WD is a microblog Douban data set. The five baseline methods are: MAUIL is a semi-supervised framework that integrates multi-granularity text features (character level, word level and topic level) and user-aligned structural information; GAlign framework aligns the network through multi-order node representation learned by graph convolution network, and combines data enhancement mechanism to process noise; CENALP is a unified framework that uses a cross-network Skip-gram model to capture network topology representation, which can jointly optimize network alignment and link prediction; Grad-Align is a graph neural network method that uses Tversky index to measure the non-symmetric set similarity between node representations, thereby realizing user identity alignment; NeXtAlign adjusts the relative position after each iteration in the attention-based architecture through the method of random walk with restart, thereby incorporating the relative position information of anchor nodes; NetTrans is based on an encoder-decoder architecture that can learn nonlinear transformation to capture the structure and attribute information between networks, thereby aligning. The prediction result can be applied to various practical scenarios, such as cross-network recommendation, user portrait construction and malicious entity detection, etc., providing support for related service development.

Claims

1. A user identity association method for mitigating cross-platform heterogeneity, the method comprising: By the semantic perception representation alignment module and the topology perception representation alignment module, combined with the adaptive integration mechanism of dynamically weighted text attribute representation and topology attribute representation, cross-platform association of user identity is realized, specifically including the following steps: S1 node representation learning: feature learning is performed on the text attributes generated by the user and the topology attributes of the social network respectively; S2 consistency-enhanced learning: the extracted text attribute representation and topology attribute representation are aligned by the semantic perception representation alignment module and the topology perception representation alignment module, the consistency of the representation of the same user in different networks is enhanced, and the enhanced text content representation and topology content representation are obtained; S3: the enhanced text content representation and topology content representation obtained in S2 are input into the adaptive representation integration type, and a weighted expert output node representation is output by the expert network and the router after dynamic weighting; The training loss of the adaptive representation integration model is represented as wherein and are the final node representations of the anchor node pairs in different networks, respectively, and L is the distance between the anchor node pairs in different networks. When L ARI is the smallest, the distance between the anchor node pairs in different networks is the closest, and thus the anchor node-related representation is further optimized. S4 Anchor node link prediction: Identify users across networks by distance calculation, for the pair of anchor nodes from S3, the final representation is and Module calculates the Euclidean distance as 2. The user identity association method for mitigating cross-platform heterogeneity according to claim 1, wherein, The step S1 generates initial representations respectively using a vector generation tool and a large-scale information network embedding method, representing text attribute representations of known anchor nodes and social network topology attribute representations 3. The user identity association method for mitigating cross-platform heterogeneity according to claim 1, wherein, The step S2 is specifically as follows: S2-1 inputs the text attribute representation into a semantic-aware representation alignment module, according to anchor nodes in the alignment problem, each anchor node pair is regarded as a positive sample, and the text attribute representation of the neighbor node is regarded as a negative sample, and the loss function L attribute optimization, and finally generates an enhanced text attribute representation; S2-2 is the anchor node confirmed as the same user in different networks, and the neighborhood connection is supplemented by the bidirectional densification mechanism. The topology attribute representation after densification is input into the topology-aware representation alignment module. The same positive and negative sample selection strategy as S2-1 is used, and the loss function L topology Optimization, get enhanced topology attribute representation; The positive and negative sample selection strategies of the semantic perception module and the topology perception module are consistent, but the parameters are independent of each other; S2-3: the semantic perception representation alignment module aims to distinguish different users while retaining specific attribute features of the same user; the topology perception representation alignment module ensures that the node maintains similar structural patterns in different networks; The total loss function L of the joint training of the two modules total = λ attribute · L attribute + λ topology · L topology , the optimization objective of the integrated semantic and structural features, wherein L attribute and L topology are the loss functions of the semantic-aware representation alignment module and the topology-aware representation alignment module respectively, λ attribute and λ topology are the weighting coefficients of the two loss functions.

4. The user identity association method for mitigating cross-platform heterogeneity according to claim 1, wherein, The step S3 is specifically as follows: the adaptive representation integration model consists of two key parts: multiple expert networks and a router mechanism; Each expert network is a feedforward neural network, which independently processes the input text attribute representation and topology attribute representation respectively; The router dynamically calculates the weight according to the input representation to determine the relevance of each expert network.

Images