An approach for assessing similarity metrics used in metric-based clone detection techniques

Similarity is an important concept in information theory. A challenging question is how to measure the amount of shared information between two systems. A large number of metrics are proposed and used to measure similarity between two computer programs or two portions of the same program. In this paper, we present an approach for assessing which metrics are most useful for similarity prediction in the context of clone detection. The presented approach uses clustering to identify clone candidates. In the experiments conducted, we applied sequential clustering using all possible permutations of a subset of the metrics used in metric-based clone detection literature. Precision and recall are calculated in every experiment. Experimental results show that the order of the metrics used affects the results dramatically. This shows that the used metrics are of variable relevance.