Top-down program comprehension with multi-layer clustering based on LDA

Software change is a fundamental ingredient of software maintenance and evolution. During software maintenance and evolution, developers usually need to understand the system quickly and accurately. With the increasing size and complexity of the evolving system, program comprehension becomes an increasingly difficult activity. In this paper, we propose a novel top-down program comprehension approach which utilizes the Latent Dirichlet Allocation (LDA) model to cluster the whole system from coarse class level to finer method level. Our approach provides a multi-layer view of the whole system at different granularity levels and supports a stepwise comprehension activity, which can effectively guide developers to quickly understand the whole system. This paper outlines the details of how to cluster the system in multiple layers based on LDA, and describes the evaluation plans.