关键词: 分类算法, 线性模型, 错误定位, 程序谱, 软件测试

Abstract: Spectrum-Based Fault Localization（SBFL） techniques aid developers to reduce the debugging effort. As a light-weight promising approach, SBFL only collects the testing result of passed or failed, and the corresponding coverage information. Based on these data, SBFL can then calculate a runtime spectra for each program statement. SBFL approaches apply various functions to map four profile features to a suspiciousness score. However, existing functions don’t give good accuracy due to the influence of the fixed parameters. Therefore, a machine learning method is proposed that can automatically construct a suspiciousness function of the specific program set. First, the old versions of a program having fault code are collected. Next, it is mapped from the feature difference in a pair of faulty statement and non-faulty statement to an instance in training dataset. Finally the linear classification algorithm of Weka is applied to learn a mapping function. The function learned from old versions is defined as the fault localization model of the program. To assess the validity of the proposed method, an experiment is performed on three benchmark datasets: Siemens suite, space and gzip. Experimental result demonstrates that the proposed method reduces fault localization cost that exists in SBFL approaches.

Key words: classification algorithm, linear model, fault localization, program spectra, software testing