Revisiting Precision and Recall Definition for Generative Model Evaluation

TL;DR

This paper redefines Precision-Recall curves for generative models, extending their applicability, linking them to classifier error rates, and proposing a new approximation algorithm demonstrated on multi-modal datasets.

Contribution

It generalizes PR curve definitions for arbitrary measures, connects them to classifier error rates, and introduces a novel approximation algorithm for better evaluation.

Findings

PR curves can distinguish mode-collapse and quality issues.

The new formulation applies to arbitrary measures beyond finite support.

The proposed algorithm effectively estimates PR curves on complex datasets.

Abstract

In this article we revisit the definition of Precision-Recall (PR) curves for generative models proposed by Sajjadi et al. (arXiv:1806.00035). Rather than providing a scalar for generative quality, PR curves distinguish mode-collapse (poor recall) and bad quality (poor precision). We first generalize their formulation to arbitrary measures, hence removing any restriction to finite support. We also expose a bridge between PR curves and type I and type II error rates of likelihood ratio classifiers on the task of discriminating between samples of the two distributions. Building upon this new perspective, we propose a novel algorithm to approximate precision-recall curves, that shares some interesting methodological properties with the hypothesis testing technique from Lopez-Paz et al (arXiv:1610.06545). We demonstrate the interest of the proposed formulation over the original approach on controlled multi-modal datasets.