Mitigating Algorithmic and Demographic Bias in Sentiment Analysis Models Trained on E-Commerce Customer Service Social Media Threads

• Abey Litty

  Texas University

  Author

influence brand reputation management, customer retention strategies, and service quality assessment. However, these models exhibit systematic performance disparities across demographic groups due to imbalances in training data, sociolinguistic variations, and the propagation of historical biases embedded in social media language patterns. This study addresses the critical research gap in intersectional demographic bias mitigation within sentiment classification models trained on customer service social media interactions. Using a dataset of 6,538 customer service conversations extracted from Twitter interactions with @AmazonHelp, this research develops and validates a hybrid debiasing framework combining VADER lexicon-based analysis with RoBERTa transformer models, enhanced by demographic-aware T5-base data augmentation. The proposed framework achieves 89.4% classification accuracy while reducing performance gaps across gender and racial subgroups by up to 70%, particularly improving F-measure for underrepresented demographic intersections from 0.62 to 0.84. Key findings demonstrate that culturally-conscious fine-tuning and balanced training approaches significantly enhance both model fairness and overall predictive performance. The framework provides a replicable methodology for e-commerce platforms seeking to implement equitable sentiment analysis systems that accurately capture customer sentiment across diverse demographic populations, with direct implications for service quality monitoring and customer engagement strategy development.

• pdf

Copyright (c) 2026 Abey Litty (Author)

This work is licensed under a Creative Commons Attribution 4.0 International License.