Comparison of classifiers in challenge scheme

TL;DR

This paper evaluates the performance of machine learning classifiers in challenge schemes, using resampling techniques to determine the significance of small performance differences among competitors.

Contribution

It introduces a method to assess the significance of classifier performance differences in challenge settings using bootstrap resampling.

Findings

Bootstrap resampling supports decision-making in challenge evaluations.

Small performance differences among classifiers may not be statistically significant.

The approach aids in fairer comparison of algorithms in constrained challenge environments.

Abstract

In recent decades, challenges have become very popular in scientific research as these are crowdsourcing schemes. In particular, challenges are essential for developing machine learning algorithms. For the challenges settings, it is vital to establish the scientific question, the dataset (with adequate quality, quantity, diversity, and complexity), performance metrics, as well as a way to authenticate the participants' results (Gold Standard). This paper addresses the problem of evaluating the performance of different competitors (algorithms) under the restrictions imposed by the challenge scheme, such as the comparison of multiple competitors with a unique dataset (with fixed size), a minimal number of submissions and, a set of metrics chosen to assess performance. The algorithms are sorted according to the performance metric. Still, it is common to observe performance differences among competitors as small as hundredths or even thousandths, so the question is whether the differences are significant. This paper analyzes the results of the MeOffendEs@IberLEF 2021 competition and proposes to make inference through resampling techniques (bootstrap) to support Challenge organizers' decision-making.