Null hypothesis significance testing is routinely used for comparing the performance of machine learning algorithms. Here, we provide a detailed account of the major underrated problems that this common practice entails. For example, omnibus tests, such as the widely used Friedman test, are not appropriate for the comparison of multiple classifiers over diverse data sets. In contrast to the view that significance tests are essential to a sound and objective interpretation of classification results, our study suggests that no such tests are needed. Instead, greater emphasis should be placed on the magnitude of the performance difference and the investigator’s informed judgment. As an effective tool for this purpose, we propose confidence curves, which depict nested confidence intervals at all levels for the performance difference. These curves enable us to assess the compatibility of an infinite number of null hypotheses with the experimental results. We benchmarked several classifiers on multiple data sets and analyzed the results with both significance tests and confidence curves. Our conclusion is that confidence curves effectively summarize the key information needed for a meaningful interpretation of classification results while avoiding the intrinsic pitfalls of significance tests.
ASJC Scopus subject areas