Comparing published multi-label classifier performance measures to the ones obtained by a simple multi-label baseline classifier

TL;DR

This paper demonstrates that many published multi-label classifiers perform worse than a simple baseline, highlighting the need for the community to compare against such baselines and provide explanations for poor results.

Contribution

It introduces General_B, a simple multi-label baseline classifier, and shows many published results are comparable or worse, urging better benchmarking and explanation practices.

Findings

Many published results are worse than or equal to General_B

Up to 43% of results on one dataset are comparable to the baseline

Most studies lack explanations for poor classifier performance

Abstract

In supervised learning, simple baseline classifiers can be constructed by only looking at the class, i.e., ignoring any other information from the dataset. The single-label learning community frequently uses as a reference the one which always predicts the majority class. Although a classifier might perform worse than this simple baseline classifier, this behaviour requires a special explanation. Aiming to motivate the community to compare experimental results with the ones provided by a multi-label baseline classifier, calling the attention about the need of special explanations related to classifiers which perform worse than the baseline, in this work we propose the use of General_B, a multi-label baseline classifier. General_B was evaluated in contrast to results published in the literature which were carefully selected using a systematic review process. It was found that a considerable number of published results on 10 frequently used datasets are worse than or equal to the ones obtained by General_B, and for one dataset it reaches up to 43% of the dataset published results. Moreover, although a simple baseline classifier was not considered in these publications, it was observed that even for very poor results no special explanations were provided in most of them. We hope that the findings of this work would encourage the multi-label community to consider the idea of using a simple baseline classifier, such that further explanations are provided when a classifiers performs worse than a baseline.