Heterogeneous Cross-Project Defect Prediction via Optimal Transport

Abstract

Heterogeneous cross-project defect prediction (HCPDP) aims to learn a prediction model from a heterogeneous source project and then apply the model to a target project. Existing HCPDP works mapped the data of the source and target projects in a common space. However, the pre-defined forms of mapping methods often limit prediction performance and it is difficult to measure the distance between two data instances from different feature spaces. This paper introduced optimal transport (OT) theory for the first time to build the relationship between source and target data distributions, and two prediction algorithms were proposed based on OT theory. In particular, an algorithm based on the entropic Gromov-Wasserstein (EGW) discrepancy was developed to perform the HCPDP model. The proposed EGW model measures the distance between two metric spaces by learning an optimal transfer matrix with the minimum data transfer cost and avoids measuring the distance of two instances of different feature spaces. Then, to improve EGW performance, an EGW+ transport algorithm based on EGW was developed by integrating target labels. Experimental results showed the effectiveness of EGW and EGW+ methods, and proved that our methods can support developers to find the defects in the early phase of software development.

Heterogeneous Cross-Project Defect Prediction Using Encoder Networks and Transfer Learning

“…There are many HCPDP models proposed by the researchers. For the comparison, we have selected some models such as EGW [19], HDP_KS [21], CTKCCA [20], and EMKCA [18], which have been frequently used for comparison in the literature. EGW applied optimal transport theory, HDP_KS effectively aligned diverse metrics across several projects, CTKCCA utilized transfer kernel canonical correlation analysis to minimize the gap between source and target dataset, and EMKCA used multiple kernel-based learning techniques that improve the separability of historical defect data by mapping it into a high-dimensional feature space.…”

“…Zong et al [19] propose a novel method for HCPDP based on optimal transport. The proposed method functions by first learning mapping from the target project features and source project features.…”

Multi-Source Heterogeneous Kernel Mapping in Software Defect Prediction

“…Zong et al [23] recently introduced optimal transport (OT) theory, and proposed two prediction algorithms for HDP using optimal transport: one algorithm is based on the entropic Gromov-Wasserstein (EGW) discrepancy, while the other is the EGW+ transport algorithm. Specifically, the method first computes a transportation plan that minimizes the discrepancy between the source and target distributions.…”

“…For this situation, traditional CPDP methods are not applicable and have limitations. To solve the inconsistency of metrics between source and target projects, researchers have proposed heterogeneous defect prediction (HDP) approaches [9][10][11][12][13][14][15][16][17][18][19][20][21][22][23].…”