Correlation-Guided Graph Transformer Networks for Cold-Start Group Event Recommendation in Social Networks

The emergence of social platforms has seen the importance of group event recommendation increase as a result of the fast rate of growth. In comparison to single-user recommendation, the group recommendation has to capture overall preferences and overcome the extreme sparsity, long-tail distributions, and cold-start problems. This paper suggests Correlation-Guided Graph Transformer Network (CG-GTN) to address these problems. The framework proposes the use of correlation to construct bi-partite graphs by using a convex combination of behavioral similarity, contextual similarity and time proximity which allows meaningful relationships to develop despite limited contacts. It uses a multi-head Graph Transformer encoder to preserve multi-hop neighborhood high-order dependencies, and an attention-based aggregation module to encode collective dynamics of decision to produce expressive group embeddings. An objective that optimizes Top-K event recommendations is based on ranking. Experiments on the Nashville Meetup dataset (where missing interactions are approximately more than 99.8) are highly effective (Attaining Precision@20 = 0.0069, Recall@20 = 0.0495 and NDCG@20 = 0.0253), which achieves the best performance among all evaluated methods, including the strongest baseline, HGT. Convergence analysis also represents quicker and more consistent training. The presented results identify the importance of correlation-directed graph modelling and transformer-based representation learning in enhancing their robustness and ranking.