Hybrid Framework for Software Change Impact Analysis Using Data-Flow Analysis and Optimised Latent Dirichlet Allocation

Abstract

Software systems continually evolve, making change management a vital yet costly part of software lifecycle processes. Change Impact Analysis (CIA) is essential to foresee the effects of modifications, minimizing unintended side effects and managing ripple impacts. Existing CIA techniques, however, often struggle to balance precision and recall, impacting their effectiveness. Many rely exclusively on either structured or unstructured source code data, missing out on insights offered by the other, which limits impact prediction accuracy. To address this, a hybrid framework for static source code CIA has been proposed, combining analyses of both structured and unstructured code information at the method level. This framework employs data-flow analysis for structured data and an optimized Latent Dirichlet Allocation (LDA) model for unstructured data. The LDA hyperparameters are fine-tuned using a differential evolution algorithm to approach optimal settings. The resulting impact sets from both analyses are merged through a union operation to capture comprehensive change effects. Experimental evaluation on three open-source Java projects demonstrated that this hybrid approach doubled the performance compared to using structural analysis alone, while preserving comparable precision and recall. Consequently, this framework offers the potential to reduce maintenance costs and optimize resource allocation. Future work may explore alternative LDA variants and optimization techniques to further enhance CIA accuracy.