The rapid expansion of social media platforms has intensified the spread of negative sentiments, misinformation, and hostile digital interactions, producing measurable consequences for public discourse, institutional trust, and societal stability. Existing models often overlook the cognitive and network-structural mechanisms that drive negative sentiment cascades, limiting capacity for early detection and effective intervention. The proposed study introduces the Expression Capacity-based Negative Sentiment (ECNS) Mitigation framework that integrates graph-theoretic analysis with psychological modeling to better understand and mitigate sentiment propagation. The ECNS architecture consists of five interconnected models dedicated to data ingestion, influencer identification, cascade monitoring, intervention control, and final sentiment-state generation. The operational core relies on two algorithms: the first identifies high-impact influencer nodes using community-level analysis and expression-capacity thresholds, while the second monitors negativity diffusion, evaluates sequential comment behavior, and applies targeted mitigation based on capacity depletion and sentiment intensity. This design enables proactive control of sentiment cascades rather than reactive moderation. The framework was evaluated across three heterogeneous datasets ResearchGate, Zhihu, and Sentiment140 to reflect diverse interaction patterns and topical domains. Comparative performance against three leading models such as EANN, HMCan, and AOAN demonstrates that ECNS provides more accurate influencer detection, stronger cascade containment, and more effective sentiment reduction. Overall, ECNS achieved improvements ranging from 5.8% to 11.4% points across key evaluation metrics, confirming its capacity to suppress negative sentiment propagation with significantly higher reliability.