Investigating the Impact of Metric Aggregation Techniques on Defect Prediction

TL;DR

This study evaluates how different metric aggregation techniques affect defect prediction models, finding that simple summation generally outperforms more complex methods and combining multiple techniques offers no significant benefit.

Contribution

The paper systematically compares nine aggregation techniques for code metrics and demonstrates that summation remains the most effective method for defect prediction.

Findings

Summation aggregation yields better defect prediction performance.

Complex aggregation methods do not significantly outperform simple summation.

Combining multiple aggregation techniques does not improve prediction accuracy.

Abstract

Code metrics collected at the method level are often aggregated using summation to capture system properties at higher levels (e.g., file- or package-level). Since defect data is often available at these higher levels, this aggregation allows researchers to build defect prediction models. Recent findings by Landman et al. indicate that aggregation is likely to inflate the correlation between size and complexity metrics. In this paper, we explore the effect of nine aggregation techniques on the correlation between three types of code metrics, namely Lines of Code, McCabe, and Halstead metrics. In addition to summation, we study aggregation techniques that are measures of: (1) central tendency (average and median), (2) dispersion (standard deviation and inter-quartile range), (3) shape (skewness and kurtosis), and (4) income inequality (Theil index and Gini coefficient). Our results show that defect prediction models built using summation outperform those built using other aggregation techniques. We also find that more complex aggregations are no different than much simpler ones and that incorporating all aggregation types in the same model does not provide a significant improvement over using summation alone.