Leveraging High‐Impact User Interactions for Social User Opinion Prediction: An Approach via Block‐Wise Matrix Factorisation and Attention Mechanism

ABSTRACT

Rapidly growing social media has become a key platform in recent years for influencing public sentiment and shaping hot public opinion. Against this backdrop, accurately predicting social users' opinions holds significant importance for public opinion analysis and guidance. Existing research generally focuses on the macro level of hot events. However, it fails to fully exploit the interaction data between high‐impact users and ordinary users. Moreover, such studies often neglect individual user characteristics and behavioural differences, which limits their capability in accurately forecasting user opinion trends. This paper proposes a novel approach to predict social users' opinions by leveraging high‐impact users' interaction data. The approach utilises the posts and interactive comments of high‐impact users to construct multiple user–post opinion matrices and employs block‐wise matrix factorisation to extract the latent embeddings of users and posts. Building upon this, the historical comment–post pairs are encoded using BERT, and a bidirectional cross‐attention mechanism is applied to model the semantic correlations between comments and posts, yielding cross‐attentional pair representations. To integrate these representations, a dynamic attention mechanism is used to weight and aggregate historical behaviours, generating an attention‐weighted semantic representation that is highly relevant to the current task. Finally, the multi‐source features of users and posts are fused through a transformer architecture and fed into a multi‐layer perceptron for opinion prediction. Experiments on a real‐world dataset demonstrate that the proposed approach achieves superior performance in user opinion prediction, showing promising potential for practical applications.