A software defect prediction method enhanced by extraction and fusion of traditional metric features and semantic features

Abstract With the advancement of software defect prediction techniques, traditional metric features and semantic features have become two primary types of features used in constructing software defect prediction methods. Traditional metric features often fail to capture program semantics, while semantic features cannot represent statistical indicators of source code. Thus, relying solely on one type of feature poses certain limitations in building defect prediction models. To fully leverage the advantages of both feature types, we propose a software defect prediction method capable of fusing metric features and semantic features (Semn-CBL). Specifically, we employ a CNN-BiLSTM hybrid model to enhance the extraction of semantic features from source code. Subsequently, we use our designed EMLP model to perform dimension alignment and feature enhancement transformation on the high-dimensional semantic features and low-dimensional traditional features. Finally, a gating mechanism is applied to fuse the semantically enhanced features (transformed by EMLP) and the traditional features for defect prediction. Experiments conducted on nine open-source projects demonstrate that the Semn-CBL model outperforms six state-of-the-art baseline methods, achieving average improvements of 14.5% in F1, 7.7% in AUC, and 10.3% in G-mean.