Improving Software Quality Prediction by Noise Filtering Techniques

Accuracy of machine learners is affected by quality of the data thelearners are induced on. In this paper, quality of the training datasetis improved by removing instances detected as noisy by the PartitioningFilter. The fit dataset is first split into subsets, and different baselearners are induced on each of these splits. The predictions arecombined in such a way that an instance is identified as noisy if it ismisclassified by a certain number of base learners. Two versions of thePartitioning Filter are used: Multiple-Partitioning Filter andIterative-Partitioning Filter. The number of instances removed by thefilters is tuned by the voting scheme of the filter and the number ofiterations. The primary aim of this study is to compare the predictiveperformances of the final models built on the filtered and theun-filtered training datasets. A case study of software measurement dataof a high assurance software project is performed. It is shown thatpredictive performances of models built on the filtered fit datasets andevaluated on a noisy test dataset are generally better than those builton the noisy (un-filtered) fit dataset. However, predictive performancebased on certain aggressive filters is affected by presence of noise inthe evaluation dataset.