A Novel Rank Aggregation-Based Hybrid Multifilter Wrapper Feature Selection Method in Software Defect Prediction.

Area of Science:

• Software Engineering
• Machine Learning
• Data Mining

Background:

• High dimensionality of software metrics poses a challenge for software defect prediction (SDP) models.
• Existing hybrid feature selection (HFS) methods inherit limitations from filter and wrapper approaches.
• Selecting optimal filter methods for HFS and mitigating wrapper inefficiencies remain critical issues.

Purpose of the Study:

• To propose a novel rank aggregation-based hybrid multifilter wrapper feature selection (RAHMFWFS) method.
• To address the filter rank selection problem and local optima stagnation in HFS.
• To enhance the performance and efficiency of software defect prediction models.

Main Methods:

• Developed a two-stage RAHMFWFS method: rank aggregation-based multifilter feature selection (RMFFS) and enhanced wrapper feature selection (EWFS).
• RMFFS aggregates multiple filter method ranks to create a robust feature list.
• EWFS uses a dynamic reranking strategy to optimize feature subset selection, reducing evaluations.

Main Results:

• RAHMFWFS effectively addressed filter rank selection and local optima stagnation issues in HFS.
• The method successfully selected optimal and non-redundant features from software defect datasets.
• Experimental results demonstrated maintained or enhanced prediction performance of SDP models (accuracy, AUC, F-measure) compared to existing HFS methods.

Conclusions:

• The proposed RAHMFWFS method is effective for improving software defect prediction.
• RAHMFWFS offers a robust solution for feature selection challenges in SDP.
• This approach enhances model performance and efficiency, outperforming current state-of-the-art HFS methods.