GICDM: Mitigating Hubness for Reliable Distance-Based Generative Model Evaluation

TL;DR

This paper introduces GICDM, a method to correct hubness in high-dimensional embedding spaces, improving the reliability of distance-based generative model evaluation and aligning better with human judgment.

Contribution

GICDM extends ICDM with a multi-scale approach to mitigate hubness effects, enhancing the accuracy of distance metrics for evaluating generative models.

Findings

GICDM effectively resolves hubness-induced failures.

It restores reliable metric behavior in high-dimensional spaces.

GICDM improves alignment with human judgment in evaluations.

Abstract

Generative model evaluation commonly relies on high-dimensional embedding spaces to compute distances between samples. We show that dataset representations in these spaces are affected by the hubness phenomenon, which distorts nearest neighbor relationships and biases distance-based metrics. Building on the classical Iterative Contextual Dissimilarity Measure (ICDM), we introduce Generative ICDM (GICDM), a method to correct neighborhood estimation for both real and generated data. We introduce a multi-scale extension to improve empirical behavior. Extensive experiments on synthetic and real benchmarks demonstrate that GICDM resolves hubness-induced failures, restores reliable metric behavior, and improves alignment with human judgment.