SpheriQ: Probabilistic Hyperbolic Reasoning for Interpretable Recommendation

ICLR 2026 Conference Submission20617 Authors

Keywords: Probabilistic recommendation, Hyperbolic embeddings, Uncertainty calibration, Explainable recommendation, Semantic reasoning

TL;DR: SpheriQ models users, items, and tags as probabilistic spheres in hyperbolic space, enabling calibrated, interpretable, and semantically diverse recommendations

Abstract: Recommendation models are increasingly used in settings where ambiguity and transparency matter, yet many approaches are deterministic or poorly calibrated. We present SpheriQ, a geometric framework that embeds users, items, and tags as probabilistic regions in hyperbolic space. The center encodes the semantic position while the radii capture the predictive uncertainty; a Gaussian semantic kernel on the manifold enables a calibrated, transitive composition along the user--tag--item paths. This bridges symbolic and distance-based paradigms, providing concept-level traces and confidence estimates with the efficiency of lightweight embeddings. We instantiate SpheriQ with automatic tag construction and Riemannian optimisation, and evaluate it on news, books, and commonsense reasoning benchmarks. Across datasets, our model pairs strong ranking performance with improved calibration and semantic diversity, while remaining training efficient. These results indicate that probabilistic geometry combined with concept-level reasoning is a practical route to trustworthy recommendation.

Primary Area: learning on graphs and other geometries & topologies

Submission Number: 20617