Understanding critical-class prioritization in class integration test order generation with a GNN-D3QN framework
TL;DR
Investigation of when and why critical-class-first scheduling is effective in CITO and a GNN-D3QN joint framework is proposed to unify critical class identification and test order optimization suggests that the effectiveness of critical-class prioritization is structurally conditioned by the topological roles and distribution patterns of critical classes.
Abstract
Although critical-class prioritization has been introduced into Class Integration Test Order (CITO) generation, its effect on test stubbing complexity remains insufficiently understood. This study aims to investigate when and why critical-class-first scheduling is effective in CITO. A GNN-D3QN joint framework is proposed to unify critical class identification and test order optimization. Specifically, a graph neural network is employed to identify critical classes from software dependency graphs, and the resulting criticality scores are incorporated into the reward function of D3QN together with test stubbing complexity. A parameter λ\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\lambda $$\end{document} is introduced to balance critical-class prioritization and stub-cost minimization. Experiments show that the proposed method can effectively identify critical classes and significantly change their positions in the generated test order. More importantly, the impact of critical-class prioritization on test cost is system-dependent: it reduces test cost in two systems, remains nearly neutral within a low-λ\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\lambda $$\end{document} range in three systems, and degrades it in four systems. Further analysis suggests that the effectiveness of critical-class prioritization is structurally conditioned by the topological roles and distribution patterns of critical classes. It tends to reduce test cost when critical classes are concentrated in high in-degree, low out-degree positions or around a few upstream super-hubs, but may become counterproductive when critical classes simultaneously span upstream and downstream roles or are topologically dispersed. Critical-class-first scheduling is not a universally effective strategy in CITO, but a conditional optimization mechanism constrained by software dependency structure. In addition, it can provide search guidance for reinforcement learning by alleviating long-range credit assignment difficulty.
