Research On Composite Attribute Selection Algorithm Catered To Software Metric

With rapid development of software process management, prediction of software reliability becomes very crucial in this process while accuracy of the prediction is highly dependent on selection of metric attributes. Due to inconsistent standard of software metric attributes, complicated system structure and redundant functional points, there are three major challenges that needs to be overcome when handling software metric data. First of all, source data needs to be pre-processed to prohibit irrelevant and redundant attributes entering the selection process. Secondly, a suitable searching strategy needs to be proposed to generate multiple feature subsets from the numerous metric attributes set without omitting one single feature. Last but not least, a practical assessment strategy which is capable to analyze the feature subset and find optimal subset is essential.One of the major objective of this thesis is to analyze the feasibility of traditional algorithms in the process of data pre-processment, feature subset assessment and searching strategy with demand and characteristics of feature selection in software metric. Moreover, based on Relief, GA and backtracking, a RGB hybrid feature selection algorithm is proposed as an effective approach to optimize current feature selection algorithms. The main work of this thesis is as follows:1) Based on the three challenges of metric attributes selection, feasibility of different algorithm is analyzed with combination of their characteristics. Moreover, the most suitable algorithms are selected by analysis and comparison in data pre-processing, feature assessment and searching strategy, with respect to performance, accuracy, efficiency etc.2) An composite attribute selection algorithm(RGB) is proposed to resolve difficulties when selecting software metric attribute based on Relief, GA and backtracking. Attributes that are irrelevant to classification is excluded using Relief. The initial population of genetic algorithm with a prior feature sequences is then provided according to the assessment of each feature. At last, optimal feature subset from multiple feature subsets was obtained by generic algorithm and backtracking.3) RGB, whose performance is tested using the NASA software metric project data, is compared with several other traditional feature selection algorithms with respect to classification accuracy, running time of the algorithm and size of feature subset. Then the experiment results are further analyzed to find out the possible mechanism and explanation of this reaults. Current work shows that RGB algorithm in software metric attributes selection preforms better than other absolute feature selection algorithm.RGB algorithm proposed in this thesis has good performance when selecting attributes in software metric. With higher classification accuracy rate and acceptable operating efficiency and feature subset size, this algorithm can effectively solve difficulties such as complicated software metric attributes and redundant metric data compared with traditional algorithms. With high probability to obtain the optimal feature subset, RGB algorithm can precisely predict the reliability of software and provide a concrete guarantee for the software development.