Community-Aware Two-Stage Diversification for Social Media User Recommendation with Graph Neural Networks

The occurrence of filter bubbles and echo chambers in social media recommendation systems poses a significant threat to information diversity and democratic discourse. Although graph neural networks (GNNs) achieve leading accuracy in user recommendation, their optimization for engagement metrics inadvertently reinforces homophily, creating isolated information ecosystems. This research developed community-aware two-stage diversification with GNNs (CATD-GNN), a method that leverages the inherent community structure of social networks to promote diversity without sacrificing recommendation quality. CATD-GNN integrates community detection with GNN learning through a two-stage diversification process. The proposed method employs the Louvain method to identify community structures as pseudo-categories, then applies submodular neighbor selection and community-based loss reweighting during GNN training (Stage 1), followed by coverage and redundancy-aware reranking (Stage 2). Twitter data capturing Black Lives Matter discourse and Reddit political discussion networks were used to evaluate the method. CATD-GNN achieves improvements in diversity metrics while maintaining competitive accuracy. The two-stage architecture demonstrates a synergistic effect: the combination of diversity-aware training and coverage-based reranking produces greater improvements than either component alone. The proposed method successfully identifies and recommends users from different communities while preserving recommendation relevance.