Sentiment on social media is dynamic and context-sensitive, often shifting over time among various emotional or evaluative states. In this work, we introduce a multi-state Cox regression framework to model time-dependent transitions in sentiment shifts. Unlike traditional sentiment analysis methods, our approach incorporates transition-specific hazard functions with time-varying covariates derived from an evolving graph-based structure that learns novel, discriminative descriptors of sentiment beyond standard sentiment indicators. Each transition between sentiment states is associated with unique baseline hazards and regression coefficients, enabling precise modeling of state-specific effects. Leveraging the Cox model’s ability to handle time-dependent transition matrices, our framework robustly predicts both the timing and direction of sentiment shifts. Tested on social media datasets spanning multiple contexts, our approach captures the complex temporal dynamics of sentiment evolution.