Avoid Oversimplifications in Machine Learning: Going beyond the Class-Prediction Accuracy

摘要

Class-prediction accuracy provides a quick but superficial way of determining classifier performance. It does not inform on the reproducibility of the findings or whether the selected or constructed features used are meaningful and specific. Furthermore, the class-prediction accuracy oversummarizes and does not inform on how training and learning have been accomplished: two classifiers providing the same performance in one validation can disagree on many future validations. It does not provide explainability in its decision-making process and is not objective, as its value is also affected by class proportions in the validation set. Despite these issues, this does not mean we should omit the class-prediction accuracy. Instead, it needs to be enriched with accompanying evidence and tests that supplement and contextualize the reported accuracy. This additional evidence serves as augmentations and can help us perform machine learning better while avoiding naive reliance on oversimplified metrics.