TAGRec: Tailness-Aware Gate for Sequential Recommendation under Long-Tailed Distributions

Abstract

Long-tailed interactions make collaborative evidence vary sharply across user histories. Existing dual-view methods alleviate sparsity but typically use static fusion and uniform auxiliary supervision. We propose TAGRec (Tailness-Aware Gated Recommendation), a unified framework in which TAG estimates contextual tailness from observable histories, T-MoV uses the resulting sample-level gate g(i) to balance semantic and collaborative views, and ASCL preserves the semantic view as a stable anchor while adapting contrastive supervision on the collaborative branch. ASCL consumes a stop-gradient copy of g(i), so it follows the routing decision without directly optimizing the gate to reduce its own weight. Experiments on four public benchmarks show best or near-best performance, especially in tail-heavy settings, with only lightweight overhead and no training-only branch retained at inference.