Socio_Semantic_Community_Detection__Integrating_Deep_Sentiment_Profiles_with_Directed_Interaction_Networks_in_Online_Social_Media

Abstract

Analyzing Online Social Networks (OSNs) requires capturing both the structural topological properties of user interactions and the semantic context of their discourse. Conventional analytical pipelines typically treat network topology and semantic content in isolation, yielding an incomplete representation of group dynamics. This study presents a three-stage, integrated socio-semantic framework that combines deep sentiment analysis with modularity-based community detection across multiple interaction networks. Utilizing a curated dataset of 410,727 COVID-19-related Twitter posts, interaction graphs were constructed using three structural modalities: undirected hashtag co-occurrence, undirected URL co-occurrence, and directed user mentions. For sentiment profiling, a Bidirectional Long Short-Term Memory (Bi-LSTM) network was trained on VADER pseudo-labeled data, achieving a classification accuracy of 97.55% and outperforming Convolutional Neural Networks (96.95%), Logistic Regression (93.30%), and Naive Bayes (82.46%). Integrating these sentiment profiles into network structures demonstrated that directed mention graphs yield the most cohesive conversational clusters. Applied to this topology, the Leiden algorithm achieved an optimal modularity score of Q = 0.9724, effectively resolving the resolution-limit defects inherent in the Louvain method (Q = 0.9720) and vastly outperforming Asynchronous Label Propagation (Q = 0.8994). These findings demonstrate that explicitly directed conversational ties, when enriched with deep context-aware sentiment attributes, provide a robust, mathematically sound paradigm for identifying and interpreting homophilic and polarized online discourse.