MLR-predictor: a versatile and efficient computational framework for multi-label requirements classification.

Area of Science:

• Software Engineering
• Machine Learning
• Natural Language Processing

Background:

• Requirements classification is crucial for successful software development, aiding in risk identification and milestone achievement.
• Existing machine learning models for requirements classification often struggle with multi-label data, exhibiting suboptimal predictive performance.
• There is a need for advanced predictors that can effectively handle multi-label requirements classification.

Purpose of the Study:

• To introduce and evaluate the MLR-Predictor, a novel approach for multi-label requirements classification.
• To demonstrate the superior predictive performance of MLR-Predictor compared to existing machine learning and deep learning models.
• To assess the generalizability and effectiveness of MLR-Predictor across different datasets and applications.

Main Methods:

• The MLR-Predictor utilizes the OkapiBM25 model to convert requirement texts into statistical vectors.
• It transforms multi-label classification data into a multi-class problem, employing a logistic regression classifier.
• Performance was evaluated against 123 machine learning and 9 deep learning pipelines on three benchmark datasets using eight metrics.

Main Results:

• MLR-Predictor significantly outperformed 123 machine learning and 9 deep learning pipelines, as well as state-of-the-art predictors.
• Achieved a 13% improvement in macro F1-measure on the PROMISE dataset compared to the state-of-the-art.
• Demonstrated 1% and 2.5% improvements on EHR-binary and EHR-multiclass datasets, respectively.

Conclusions:

• The MLR-Predictor offers a robust and effective solution for requirements classification, outperforming current state-of-the-art methods.
• Its effectiveness was further validated in a case study on customer review classification, outperforming BERT by 1.4% F-1 score.
• The findings highlight MLR-Predictor's utility and generalizability for diverse requirements classification tasks.