SaFeAU framework beats GCNs on sparse recommendation data

A new paper accepted at KDD 2026 introduces SaFeAU (Semantic Factor enhanced Alignment and Uniformity), a framework that overcomes a fundamental flaw in collaborative filtering (CF). Traditional CF methods treat all non-interacted user-item pairs as negative samples, even when users would have liked those items—causing false negatives that limit model generalization. SaFeAU addresses this by first using Semantic Factor Routing to disentangle item representations into independent, global semantic factors, then applying Semantic Factor Matching to identify uninteracted items sharing those factors as potential positive pairs, enriching sparse supervision signals.

SaFeAU consists of three components: Semantic Factor Routing (SFR), Semantic Factor Matching (SFM), and Semantic Pairs Alignment (SPA). It enables matrix factorization to capture high-order collaborative signals without expensive graph neural networks, avoiding over-smoothing and computational bottlenecks. Experiments on four sparse real-world datasets show SaFeAU consistently outperforms both GCN-based and MF-based state-of-the-art methods in recommendation accuracy while being computationally more efficient. The work introduces a new semantic-enhanced learning paradigm for recommender systems.