A Study on Predicting Dynamic User Behavior in Social Media Based on Time Series Analysis and Multi-Feature Regression Models

Abstract

: This paper constructs a multi-level mathematical model to achieve precise prediction of complex user behavior within the dynamic ecosystem formed by content production and user interaction on social media platforms. In studying the evolution of new follower growth for influencers, an ARIMA time series model was constructed. Data stationarity was achieved through ADF testing and first-order differencing techniques, with model parameters optimized using the AIC criterion. This successfully predicted traffic growth trends for key influencers like B21 and B5, providing quantitative insights for platform traffic allocation. For modeling user follow intent, this paper proposes a multi-feature composite prediction framework integrating five key characteristics: weighted interaction intensity, composite behavior density, behavioral diversity, and others. By combining ARIMA, exponential smoothing, and linear trend extrapolation, it transforms complex social interactions into quantifiable predictive metrics. Furthermore, for real-time activity state identification, a dual-prediction system combining logistic regression and linear regression was established. By extracting hourly activity patterns and historical interaction matrices, it simultaneously discerns user online probability during specific time periods and high-frequency interaction targets. Research findings demonstrate that this series of models effectively captures the dynamic fluctuation patterns of social platform data. The output results for the top-3 interacting bloggers exhibit high predictive accuracy, laying a theoretical foundation for optimizing platform recommendation algorithms and enhancing user stickiness.