CAFEM: Context-Aware Feature Enhancement with Multi-Stage Sampling for Statement-Level Defect Prediction

Software defect prediction is one of the key technologies to ensure software quality. The current software quality assurance engineering needs more fine-grained defect prediction methods. In this paper, we propose a deep learning-based statement-level software defect prediction method, which integrates lexical analysis features and expert measurement features to accurately locate code defects. In terms of model architecture, a context-aware adaptive feature enhancement module (CAFEM) is proposed, which enhances feature representation capability through self-focusing mechanism and adaptive gating structure. To solve the problem of class imbalance in software defect prediction, an adaptive combination loss function based on binary crossentropy,Focal loss and F1 score was designed. SMOTE oversampling and TomekLinks undersampling were used to optimize the sample distribution. Experiments on a largescale benchmark dataset containing 119,989 code samples showed that our model establishes fine performance with an accuracy of 0.7315, a precision of 0.6364, a recall rate of 0.8941, and an F1 score of 0.7433, reaching the highest values on several metrics compared to existing methods.