Probabilistic component identification

Abstract

Automatic component identification is an important step for many software engineering tasks and tools, from reverse engineering to program comprehension. Although being a topic of research for a long time, generally applicable solutions remain elusive with very low recall. Most component identification techniques impose structure on the software, therefore requiring substantial knowledge about the software, instead of inferring such structure.We propose PCI, a non-parametric, probabilistic algorithm to the problem of automatic component identification. The proposed approach uses structural, behavioural and lexical features of the software to infer the underlying structure and identify the components. In order to evaluate this approach, multiple experiments are performed on a software system for which components have previously been identified [by others]. The proposed algorithm is compared to two existing algorithms. The experiments measure precision and recall, and attempt to unveil the impact of each source of information on the accuracy of the approach. We find PCI outperforms the others most notably on recall, with improvements that go from 4x to 10x better recall.