Evaluating software clustering algorithms in the context of program comprehension

We propose a novel approach for evaluating software clustering algorithms in the context of program comprehension. Based on the assumption that program comprehension is a task-driven activity, our approach utilizes interaction logs from previous maintenance sessions to automatically devise multiple comprehension-aware and task-sensitive decompositions of software systems. These decompositions are then used as authoritative figures to evaluate the effectiveness of various clustering algorithms. Our approach addresses several challenges associated with evaluating clustering algorithms externally using expert-driven authoritative decompositions. Such limitations include the subjectivity of human experts, the availability of such authoritative figures, and the decaying structure of software systems. We conduct an experimental analysis using two datasets, including an open-source system and a proprietary system, to test the applicability of our approach and validate our research claims.