Social Networks Link Prediction Based on Incremental Learning

Objective Various types of relationships in social networks exist between nodes, and the number of nodes changes over time. The heterogeneity and dynamic topology pose significant challenges for link prediction. Specifically, this study addresses three key issues: (1) an incremental update strategy for random walk sequences, (2) the extraction of the correlation of causal relationships, and (3) the construction of the mutual perceptron.Methods Therefore, an incremental learning social network link prediction (IL−SNLP) method is proposed. IL−SNLP consists of two components: the node embedding model and the prediction model. The node embedding model structures the network in layers based on relationship types. An incremental update strategy is designed for each network layer to generate updated random walk sequences, employing a temporal random walk approach. The incremental skip-gram model is then utilized to extract the embedding vectors of nodes from the random walk sequences in each layer. A probabilistic model is employed to extract the correlation of causal relationships between relationship types, enhancing the representation of nodes across different layers. In the prediction model, a multilayer perceptron (MLP) is utilized to construct the mutual perceptron, which predicts links by processing the embedding vectors of node pairs. The specific process of the node embedding model consists of temporal random walks, an incremental update strategy for random walk sequences, an incremental skip-gram model, and extracting the correlation of causal relationships. In the process of generating random walk sequences based on temporal random walks, the next node is selected if its timestamp is greater than the current timestamp. This approach ensures that both local structural and temporal information are preserved within these sequences. An incremental update strategy is introduced to update the random walk sequences in scenarios involving data increments. This updated strategy considers three situations: completely outdated sequences, partially outdated sequences, and the emergence of new nodes. The incremental skip-gram model is employed to extract and update the embedding vectors of nodes from the revised random walk sequences. This model automatically updates the unigram table and the noise distribution of all nodes using the updated random walk sequences without requiring access to previous data. For each relationship type, a conditional probability model is employed to calculate the probability of other relationship types occurring following a given relationship type. This model extracts the correlation of causal relationships. The correlation of causal relationships is incorporated into the node embedding vectors to generate the final embedding vectors. The prediction model utilizes an MLP to construct the mutual perceptron. The contribution of a node to a link is determined using the inner product of vectors. The contribution of a node pair to a link is then obtained by summing the individual contributions of both nodes. The computed node pair contribution is passed through a sigmoid function to determine the probability of a link forming.Results and Discussions Experiments are conducted to evaluate the prediction performance of IL−SNLP on three real-world social networks with varying sizes and sparsity levels: Super User, Math Overflow, and Ask Ubuntu. These datasets contain three relationship types: answering questions, commenting on answers, and commenting on questions. The evaluation metrics include AUC and F1−score. The experiments encompass several aspects, including comparative analysis, validation of effectiveness, assessment of the impact of the temporal random walk strategy, and evaluation of the effect of different incremental granularities on IL−SNLP. The comparison experiment results shows that the IL−SNLP enhances prediction performance over the baseline methods by 10.08% to 67.60% in AUC and 1.76% to 64.67% in F1−score. This experiment demonstrates the superior prediction performance of IL−SNLP. In addition, the baseline models do not perform effectively on sparse networks. However, the prediction performance of IL−SNLP on Ask Ubuntu and Super User decreases by only 4.54% and 5.25%, respectively, compared to its performance on Math Overflow. IL−SNLP exhibits a more stable prediction performance than the baseline methods. The validation of effectiveness experiments includes the validation of the prediction performance effectiveness of the causal relationship and mutual perceptron, as well as the validation of the time performance effectiveness of IL−SNLP. These results showed 1) that IL−SNLP improves the AUC by 51.82% to 67.78% and the F1−score by 61.47% to 81.91% across the three datasets, validating the two modules’ effectiveness and demonstrates that the correlation of the causal relationship and the mutual perceptron enhances the model’s prediction performance; 2) after only 16 to 20 iterations of training with inherited parameters, the model achieves the same performance as a model trained with non-inherited parameters after 200 iterations; 3) IL−SNLP improves time efficiency by 30.78% to 257.58%. This experiment demonstrated that IL−SNLP significantly reduces the running time and validates training effectiveness with inherited parameters. In addtion, the effect of two temporal random walk strategies on IL−SNLP shows that the random walk with bias is more effective; the effect of different incremental granularity levels on IL−SNLP reveals that the prediction performance of IL−SNLP on larger-scale networks is more affected by incremental granularity.Conclusions Accordingly, IL−SNLP exhibits superior prediction performance and significantly enhances runtime efficiency. In practical application scenarios, IL−SNLP can support the discovery and prediction of anomalous relationships in social networks, ensuring their security and stability.