BigCloneBench Considered Harmful for Machine Learning

Abstract

BigCloneBench is a well-known large-scale dataset of clones mainly targeted at the evaluation of recall of clone detection tools. It has been beneficial for research on clone detection and evaluating the performance of clone detection tools, for which it has become standard. It has also been used in machine learning approaches to clone detection or code similarity detection. However, the way BigCloneBench has been constructed makes it problematic to use as ground truth for learning code similarity. This paper highlights the features of BigCloneBench that affect the ground truth quality and discusses common misperceptions about the benchmark. For example, extending or replacing the ground truth without understanding the properties of BigCloneBench often leads to wrong assumptions which can lead to invalid results. Also, a manual investigation of a sample of Weak-Type-3/Type-4 clone pairs revealed 86% of pairs to be false positives, threatening the results of machine learning approaches using BigCloneBench. We call for a halt in using BigCloneBench as the ground truth for learning code similarity.