STELAR Tech Brief #3: Bias-aware Data Augmentation for Food-Hazard ID

Biases in machine learning for agri-food safety—such as class imbalance, language variation, and text-length inconsistency—harm model accuracy. This paper presents a structured pipeline to mitigate these issues through data assessment, cleaning (duplicate removal, missing-value handling via feature mixing), and targeted augmentation. A 10-method framework (e.g., synonym replacement, style transfer, text summarisation) generates 16 synthetic data points across five classes, significantly improving balanced accuracy. Fine-tuned BART models and ChatGPT 4.0 produce consistent, low-loss synthetic data. Techniques like padding, encoding, and feature selection enhance robustness. The approach reduces computational costs while maintaining contextual accuracy, offering a scalable, responsible AI solution for food-hazard identification. Part of the EU-funded STELAR project (Grant ID: 101070122), led by Vivek Kumar at the University of the Bundeswehr Munich.