A Study on Predicting Short-Term Interest Behaviour of Social Media Users Based on Cosine Clustering and Sliding Window Logistic Regression

Abstract

In recent years, social media platforms have profoundly influenced human social interactions and information acquisition. Precise analysis of user needs enables efficient content matching, fostering positive ecosystem cycles that ultimately enhance platform competitiveness and commercial value. Addressing the requirement to "predict which users will gain new followers on 22 July" for a specific social media platform, this paper proposes a user-blogger interaction modelling approach integrating Dynamic Bayesian Network principles with collaborative filtering mechanisms. First, a normalised user-blogger interaction matrix is constructed, utilizing cosine similarity to partition users into three interest-homogeneous clusters. Subsequently, a three-day sliding window extracts daily totals of views, likes, and comments as short-term behavioural features. Independent binary Logistic regression models with L1 regularisation are trained for each cluster, addressing positive-negative sample imbalance through oversampling. Time-series cross-validation (training: 11–17 July; validation: 18–20 July) indicates optimal performance with a three-day window, achieving adjusted R² ≥ 0.84. The final model outputs the probability of all unfollowed bloggers following each other on 22 July, alongside a Top-3 recommendation. Taking U7 as an example, the probability of them following B7 reached 85.37%. Robustness and sensitivity tests demonstrate the model's sound stability and practical value for platform recommendation scenarios. This paper excels in dynamic data modelling, drawing extensively from literature to establish an optimal model with comprehensive consideration of the problem. Having passed robustness and sensitivity tests, the model holds reference value for platform push notifications.