Comparing Multi-class Classifiers: On the Similarity of Confusion Matrices for Predictive Toxicology Applications

Abstract

Calculating the similarity of predictive models helps to characterize the models diversity and to identify relevant models from a collection of models. The relevant models are considered based on their performance, calculated using their confusion matrix. In this paper, we propose a methodology to measure the similarity for predictive models performances by comparing their confusion matrices. In this research, we focus on multi-class classifiers for toxicology applications. The performance measures of confusion matrices of multi-class classifiers are regrouped into a binary classification problem. Such approach may result in selecting multi-class classifiers with lower False Negative Rate (FNR) for example. Consequently, the methodology for model comparison based on the similarity of confusion matrices provides a working way to select models from a collection of classifiers.