Data-Driven Feature Selection on Change Metric Using Neural Network: A Comparative Study in Micro-Datasets

Handani, Felix and Siahaan, Daniel and Rochimah, Siti and Aritsugi, Masayoshi (2026) Data-Driven Feature Selection on Change Metric Using Neural Network: A Comparative Study in Micro-Datasets. In: International Conference on Smart Computing, IoT, and Machine Learning (SIML) 2026, 10-11 June 2026, Universitas Muhammadiyah Surakarta, Surakarta, Indonesia.

PDF 2026168524.pdf - Accepted Version Restricted to Repository staff only Download (426kB) | Request a copy

Abstract

Feature selection is a critical phase in developing Artificial Neural Network (ANN) models for software change prediction, particularly when dealing with micro-datasets. Change metrics, often correlated with internal attributes like Object-Oriented (OO) metrics, are sensitive to multicollinearity and stochastic noise, which can severely degrade predictive models. While statistical analysis has been employed to quantify model stability, few studies have investigated feature selection procedures specifically for ANNs trained on micro-datasets. This study evaluates the impact of five feature selection techniques, comparing static thresholding methods against adaptive and robust frameworks using k-fold cross-validation on two structurally distinct software datasets. The results indicate that for collinear environments, a severe dimension reduction method guided by a Variance Inflation Factor (VIF) threshold is mandatory to resolve weight confusion. Conversely, Kendall’s Tau demonstrates superior stability in the presence of noisy data. The study shows that appropriate preprocessing turns negative R2 scores into positive predictive capabilities by prioritizing feature dimensionality reduction over model complexity. These findings offer guidelines for deploying reliable ANNs in data-scarce software engineering contexts.

Item Type:	Conference or Workshop Item (Speech)
Subjects:	T Technology > T Technology (General)
Divisions:	Faculty of Engineering > Department of Informatic
Depositing User:	FELIX HANDANI
Date Deposited:	07 Aug 2026 08:07
Last Modified:	07 Aug 2026 08:07
URI:	http://repository.ubaya.ac.id/id/eprint/51153

Actions (login required)

View Item