PREDICTING USER ACTIVITY AND INTERACTION SEQUENCES ON SOCIAL PLATFORMS BY INTEGRATING ENSEMBLE LEARNING AND TIME SERIES DECOMPOSITION

Addressing the highly discrete and complex patterns of user behavior on social platforms, this study constructs a predictive framework combining ensemble learning with dynamic time series decomposition to achieve precise modeling of user online status and interaction behaviors. Within the user online status recognition and multi-level influencer recommendation models, the study first extracts key features such as the user's past 3-day, 7-day, and historical total active days. Utilizing a random forest classifier to capture the nonlinear correlations in user behavior, it effectively determines the online status for specific dates. Subsequently, by defining an interaction score formula incorporating time-decay weights and integrating user preferences with global popularity, the model efficiently recommends interactive bloggers, achieving a recall rate of 80.48% on the validation set. In precise temporal interaction prediction based on user segmentation and time series decomposition, this study addresses the temporal periodicity of historical behavior by introducing the Facebook Prophet model. Through automatic decomposition of trend and seasonal components, it captures user activity patterns during holidays and specific periods. To enhance prediction accuracy, the study further clusters users using the K-means algorithm and configures model hyperparameters based on behavioral feature differences. Experimental results demonstrate that this approach effectively addresses the challenge of predicting interaction intensity across time intervals under sparse data conditions, providing scientific decision support for personalized platform operations.