User Service Behavior Prediction via Spatial-Temporal Fusion: A Graph Inference Perspective

To provide more stable and precise communication services for users, accurately predicting user behavior is an essential task. However, this task faces many challenges in practice, especially how to effectively represent and model the dynamic relationships between user behaviors and how to analyze the interpretability of the model. To address these challenges, we propose a dynamic graph-based method to establish dynamic associations between user behaviors and construct a spatial-temporal dynamic graph dual-channel (STDGDC) model. The core idea of the STDGDC model is to establish a full-chain technical system encompassing “dynamic modeling, conditional fusion, and mechanism interpretation." This method characterizes complex service behavior patterns through constructing dynamic spatio-temporal graphs, while incorporating environmental factors as regulatory variables during feature fusion. The conditional fusion mechanism not only preserves predictive accuracy through dynamic parameter calibration but also enhances model interpretability. Considering the huge user base and the sensitivity of personal privacy, we focus on cellular traffic data as a macro-level manifestation of user service behavior, and conducts experimental validation on public datasets. The experimental results validate the rationality and effectiveness of the proposed STDGDC model, thereby offering fresh research insights and practical approaches for predicting user behavior in the field of communication services.