Why Comparing Single Performance Scores Does Not Allow to Draw Conclusions About Machine Learning Approaches

Developing state-of-the-art approaches for specific tasks is a major driving force in our research community. Depending on the prestige of the task, publishing it can come along with a lot of visibility. The question arises how reliable are our evaluation methodologies to compare approaches? One common methodology to identify the state-of-the-art is to partition data into a train, a development and a test set. Researchers can train and tune their approach on some part of the dataset and then select the model that worked best on the development set for a final evaluation on unseen test data. Test scores from different approaches are compared, and performance differences are tested for statistical significance. In this publication, we show that there is a high risk that a statistical significance in this type of evaluation is not due to a superior learning approach. Instead, there is a high risk that the difference is due to chance. For example for the CoNLL 2003 NER dataset we observed in up to 26% of the cases type I errors (false positives) with a threshold of p < 0.05, i.e., falsely concluding a statistically significant difference between two identical approaches. We prove that this evaluation setup is unsuitable to compare learning approaches. We formalize alternative evaluation setups based on score distributions.