Role-Aware Neural Convex Divergence Heads for Asymmetric Representation Learning

2026-07-02

AI Summary

Traditional symmetric metrics struggle to model directed relationships—such as entailment or hierarchical structures—in representation learning. This work proposes a role-aware neural convex divergence head that introduces independent linear projections for source and target representations, enabling the computation of non-negative, convex, and interpretable directed scores within the Bregman divergence space induced by Input Convex Neural Networks (ICNNs). The proposed module is directional, structurally informed, and readily pluggable into existing architectures. Empirical results demonstrate its significant superiority over standard ICNN-Bregman approaches on semantic and ontological benchmarks, achieving zero negative divergence rates and improved directional accuracy. However, it slightly underperforms specialized symmetric or hyperbolic models on large-scale fixed-feature citation prediction tasks.

Abstract

Many representation learning problems involve directed relations, such as lexical entailment, sentence entailment, ontology hierarchy, and citation links. Standard Euclidean, cosine, and Mahalanobis heads are symmetric, while generic neural scorers can model directionality but provide limited geometric structure. This paper proposes a role-aware neural convex divergence head for asymmetric representation learning. The head applies source- and target-role projections before evaluating an input-convex neural Bregman divergence, yielding a nonnegative structured score in the role-projected space. We characterize its projected-space identity, source-role convexity, directional-gap decomposition, and Hessian-based local curvature. Experiments on lexical, sentence, ontology, and directed graph benchmarks compare symmetric distances, unstructured asymmetric scorers, order/hyperbolic baselines, plain ICNN-Bregman heads, and the proposed role-aware variant. Across ten random seeds on the main semantic and ontology benchmarks, role-aware projections consistently improve directional accuracy over plain ICNN-Bregman heads while preserving zero observed negative divergence rate. The results also identify a boundary case: on large fixed-feature citation prediction, specialized symmetric or hyperbolic baselines remain stronger in ranking accuracy. Overall, the proposed head is best understood as a structured and interpretable plug-in distance module for tasks where directional relations matter.

Problem

• asymmetric representation learning
• directed relations
• convex divergence
• role-aware projections
• Bregman divergence

Innovation

• asymmetric representation learning
• neural Bregman divergence
• role-aware projection
• input-convex neural network
• directed relational modeling