Dynamic interactive behavior in social media based on multidimensional hybrid prediction models and spatio-temporal feature extraction

Abstract

This paper addresses the challenge of accurately predicting influencer evolution and deep user interaction behaviors on social media platforms by constructing a comprehensive data-driven modeling framework based on multimodal data. First, core interaction data—including views, likes, comments, and follows—undergoes one-hot encoding preprocessing. A composite weighted system integrating Autoregressive Integrated Moving Average (ARIMA) models, Backpropagation (BP) neural networks, Expert Grey Model (EGM), and multiple linear regression achieves high-precision evolutionary prediction of new follower growth for influencers through error balancing. To address the conversion of latent user follow intentions, a three-layer BP neural network model with LeakyReLU activation functions was designed. By extracting temporal behavioral feature vectors, it successfully captured the nonlinear patterns governing the transition from basic interactions to follow actions. Additionally, logistic regression was employed to probabilistically model users' social online states. Combined with K-means clustering, this enabled bidirectional feature matching between users and bloggers, precisely identifying high-value interaction relationships. Finally, for fine-grained prediction of dynamic interaction periods, this study coupled linear regression with BP neural networks to achieve quantitative inversion of interaction density across 24 consecutive hourly segments. Experimental results demonstrate the system's robust performance and generalization capabilities, providing scientific foundations for platform content optimization and personalized recommendations.