A Gated Hybrid Contrastive Collaborative Filtering Recommendation

A team of researchers from multiple institutions (including Eduardo Ferreira da Silva, Maycon dos Santos Oliveira, and others) has introduced a Gated Hybrid Contrastive Collaborative Filtering framework designed to improve top-N recommendation quality. Traditional review-aware recommenders often optimize for rating prediction accuracy, which misaligns with ranking tasks. The new architecture addresses this by injecting textual review signals into an autoencoder-based collaborative model via a layer-wise adaptive gating mechanism. This gate dynamically balances traditional collaborative embeddings with topic-based features extracted from reviews. To further sharpen the latent space, a contrastive learning module aligns semantic and collaborative representations, and the entire model is trained with a pairwise Bayesian personalized ranking (BPR) objective that explicitly separates relevant from irrelevant items.

Evaluated across five configurations—ranging from pure collaborative to gated text and contrastive variants—the framework was tested on three real-world datasets: Amazon Movies & TV, IMDb, and Rotten Tomatoes. Results showed consistent improvements in hit rate @10 and normalized discounted cumulative gain @10 over state-of-the-art review-aware baselines. The study highlights the importance of controlled semantic fusion for ranking-driven recommendation, demonstrating that simply adding review features is less effective than adaptively balancing them with collaborative signals. The paper is available on arXiv (2604.27117) and is pending registration with DataCite.