.png){kind=icon}
.png){kind=icon}
Return
Instance gravity oversampling method for software defect prediction
.png){kind=icon}
PRE
AI
DOI：10.1016/j.infsof.2024.107657.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Context: In the software defect datasets, the number of defective instances is significantly lower than that of nondefective instances. This imbalance adversely impacts the predictive performance of the model. Oversampling methods can effectively balance datasets. However, traditional oversampling methods often struggle to capture the underlying relationships between features and are prone to introducing noise during instance synthesis. Objective: Inspired by the law of gravity, we propose a novel oversampling method based on instance gravity (MOSIG). Method: This method begins by introducing a new metric, instance gravity, to measure the similarity between instances. Subsequently, feature models are constructed, and instance groups are generated. Instances that meet specific conditions based on instance gravity are then identified within different instance groups. Finally, we propose a novel method for synthesizing defective instances by assigning weights to instances according to their gravity. Results: Experimental results demonstrate that MOSIG significantly enhances the predictive performance of both the CART decision tree and Naive Bayes models across 21 publicly available software defect datasets. The experimental results are further validated using the Friedman ranking and Nemenyi post-hoc test, confirming that MOSIG is statistically significant. Conclusion: MOSIG represents a more promising oversampling method.
Keywords:
Instance gravity
Oversampling method
Feature model
Instance group
Software defect prediction
Journal
IF:
4.3
Papers:
3.7K
Citations:
7.7K