Machine learning code snippets semantic classification.

Area of Science:

• Machine Learning
• Software Engineering
• Data Science

Background:

• Program code is increasingly used for data science models.
• Annotating code snippets is crucial for model training.
• The Code4ML corpus has limited labeled data (~0.2%).

Purpose of the Study:

• To develop an automated approach for classifying code snippets from the Code4ML corpus.
• To address the scarcity of high-quality labeled data for machine learning code.
• To improve the performance of data science models trained on code.

Main Methods:

• Leveraging CodeBERT, a transformer-based model, for code snippet classification.
• Developing a specialized algorithm to separate ambiguous code snippets with multiple labels.
• Employing a data augmentation strategy to expand the labeled dataset.

Main Results:

• Achieved an F1 test score of approximately 89% for code snippet classification.
• Significantly increased the amount and quality of labeled data in the Code4ML corpus.
• Demonstrated the enhanced practicality of CodeBERT for code classification tasks.

Conclusions:

• The proposed method effectively classifies machine learning code snippets.
• Automated data augmentation improves supervised model training.
• Enriching code datasets like Code4ML is vital for advancing data science models.