Stable History, Better Rankings: Invariant Parent Selection with Geodesic Attention for Recommenders

Abstract

Sequential recommender systems increasingly rely on transformer attention to model complex dependencies in user histories, yet standard attention is typically unconstrained: any candidate item may attend to any historical interaction. This flexibility can improve accuracy, but it often amplifies spurious correlations, and degrades robustness under non-stationary user behavior. We propose GeoMaskRec, a novel geometry-aware and structure-constrained architecture that introduces a new inductive bias for impression-based recommendation. GeoMaskRec is the first recommendation framework to unify three components in a single, end-to-end model: invariant parent discovery, geodesic hyperspherical attention, and regime-switching temporal influence. We evaluate GeoMaskRec on three public benchmarks spanning distinct modalities and interaction structures: MIcrosoft News Dataset (MIND) for news recommendation, user interaction logs from Globo news portal, and Amazon Reviews data for e-commerce. Across datasets, GeoMaskRec yields strong impression-level ranking performance and produces sparse, transparent attention maps that directly expose candidate-to-history influence pathways. Across the different datasets, GeoMaskRec achieves nDCG@10 / MRR@10 of 0.571/0.521, 0.606/0.675, and 0.508/0.465, respectively. Our results highlight that combining invariant causal structure with hyperspherical geometry and regime-aware temporal adaptation provides a principled and effective foundation for robust recommendation.