While behavior prediction in social networks has attracted significant research attention, accurately modeling long-term user behavior remains a challenge owing to complex temporal dynamics and psychological influences. To tackle this issue, we introduce, ChronoGAT, a novel neural network model that integrates socio-psychological mechanisms -- including primacy effect, recency effect, and social conformity -- into a multi-head graph attention architecture for improved long-term behavior forecasting. The model dynamically captures structural and temporal dependencies, with specialized attention mechanisms designed to emphasize long-range interaction patterns and psychological influences. Evaluated on a large-scale dataset from a video-sharing platform (comprising 6,789 videos, 2.63 million users, and over 3 million relations), our approach achieves a 19.4% improvement in prediction accuracy compared to state-of-the-art baselines. Notably, quantitative analysis reveals that socio-psychological factors explain over 70% of the variability in long-term content dissemination trajectories. This work offers a computationally scalable and psychologically principled framework for effectively modeling long-term user behavior in online social systems.