PEAPOD: A Heterogeneous Defect Prediction Approach Based on Deep Density Sampling and Deep Domain Adaptation

Heterogeneous Defect Prediction (HDP) refers to using labeled instance data from source projects to predict potential defect instances for the target project. This approach addresses the feature heterogeneity problem between different datasets, aiming to improve software quality actually. Notably, existing research on HDP addresses imbalance issues by focusing solely on minority classes and neglecting the influence of majority classes. Moreover, these studies fail to regard issues of feature consistency and differences in data distribution in depth while addressing heterogeneity in the feature space. Consequently, this paper proposes a 2-stage HDP approach, PEAPOD, based on dee P d Ensity s Ample and dee P d Omain a Daptation. First, PEAPOD creates a balanced dataset by transforming data into a latent space, selecting high-quality samples by density and enhancing the defective class with synthetic samples through feature-level oversampling. Next, PEAPOD aligns matching features utilizing KS-test and the Hungarian algorithm and subsequently reduces data distribution differences via a Domain Adversarial Neural Network (DANN). Extensive and rigorous experimental results on 23 projects from 5 public datasets demonstrate PEAPOD outperforms HDP baseline methods.