Adaptive Hybrid Embedding Fusion for Interpretable User-Generated Content Clustering and Engagement Analysis

Abstract

User-generated material (UGC) from social media platforms provides important insights into public behaviour and engagement trends; however, the informal and noisy structure of social media content poses major challenges for thematic analysis. In this work, we propose a hybrid clustering framework that combines TF-IDF lexical features and Sentence-BERT (SBERT) semantic embeddings for better detection of engagement-driven themes. Pre-processing of raw text, hybrid feature vectors generation and k-means clustering were performed. The silhouette analysis of the most coherent cluster configuration reveals two dominant thematic groups, activity or event-oriented posts and emotion or mood-oriented posts. Additional analysis of engagement revealed an influence of thematic context on user engagement patterns, with activity-based posts typically receiving greater engagement. The results indicate that hybrid embeddings are effective in improving cluster quality and provide a robust methodology for investigating thematic and behavioural patterns in UGC.