GFML: Gravity function for metric learning

Abstract

Diverse machine learning algorithms rely on the distance metric to compare and aggregate the information. A metric learning algorithm that captures the relevance between two vectors plays a critical role in machine learning. Metric learning may become biased toward the major classes and not be robust to the minor ones, i.e., metric learning may be vulnerable in an imbalanced dataset. We propose a gravity function-based metric learning (GFML) that captures the relationship between vectors based on the gravity function. We formulate GFML with two terms, (1) mass of the given vectors and (2) distance between the query and key vector. Mass learns the importance of the object itself, enabling robust metric learning on imbalanced datasets. GFML is simple and scalable; therefore, it can be adopted in diverse tasks. We validate that GFML improves the recommender system and image classification.