Improving Requirements Classification with SMOTE-Tomek Preprocessing
A machine-learning study on class imbalance in requirements engineering, using TF-IDF, SMOTE-Tomek, and leakage-safe stratified cross-validation.
Abstract
This research studies requirements classification as an imbalanced multiclass text-classification problem over the PROMISE requirements dataset. The work combines TF-IDF feature extraction, SMOTE-Tomek preprocessing, and stratified K-fold cross-validation in a leakage-safe evaluation protocol, applying resampling only to training folds while keeping validation folds untouched. The results show that lightweight, interpretable machine learning models can deliver practical gains in requirements engineering: Logistic Regression reaches 76.16% +/- 2.58% accuracy, with improved F1-score and MCC, while preserving interpretability through feature-coefficient analysis.
Key Innovations
Applies SMOTE-Tomek only to the training folds inside stratified K-fold cross-validation, keeping validation folds untouched for a realistic evaluation.
Uses TF-IDF and classical machine-learning models to classify 12 PROMISE requirement types, including severely underrepresented non-functional categories.
Shows strong gains for Logistic Regression, reaching 76.16% +/- 2.58% accuracy and improved MCC, while preserving readable feature-coefficient analysis.