Heterogeneous cross-company defect prediction by unified metric representation and CCA-based transfer learning

Abstract

Cross-company defect prediction (CCDP) learns a prediction model by using training data from one or multiple projects of a source company and then applies the model to the target company data. Existing CCDP methods are based on the assumption that the data of source and target companies should have the same software metrics. However, for CCDP, the source and target company data is usually heterogeneous, namely the metrics used and the size of metric set are different in the data of two companies. We call CCDP in this scenario as heterogeneous CCDP (HCCDP) task. In this paper, we aim to provide an effective solution for HCCDP. We propose a unified metric representation (UMR) for the data of source and target companies. The UMR consists of three types of metrics, i.e., the common metrics of the source and target companies, source-company specific metrics and target-company specific metrics. To construct UMR for source company data, the target-company specific metrics are set as zeros, while for UMR of the target company data, the source-company specific metrics are set as zeros. Based on the unified metric representation, we for the first time introduce canonical correlation analysis (CCA), an effective transfer learning method, into CCDP to make the data distributions of source and target companies similar. Experiments on 14 public heterogeneous datasets from four companies indicate that: 1) for HCCDP with partially different metrics, our approach significantly outperforms state-of-the-art CCDP methods; 2) for HCCDP with totally different metrics, our approach obtains comparable prediction performances in contrast with within-project prediction results. The proposed approach is effective for HCCDP.