Authors: A. Helen Nithya, V Shanmugasundaram
Abstract: The deployment of black-box AI models to predict consumer behavior has created an inherent trade-off between predictive performance and interpretability. In this paper, we propose a framework for consumer behavior prediction based on XAI models to overcome the challenge of providing reliable, interpretable, and actionable analytics in marketing. By conducting empirical studies involving 500,000 instances of consumer behavior on datasets from e-commerce and banking sectors, we examine the comparative performance of four XAI approaches – LIME, SHAP, IG, and Decision Trees – relative to conventional black-box approaches including Random Forest, XGBoost, and Neural Networks. Our findings show that SHAP provides the best prediction accuracy (92.3%) and explains decisions made using it by exhibiting better explainability levels (92% explanation accuracy for consumer decision prediction at 80-20 split ratio between training and testing sets). Product category features were found to be predominant in making predictions about purchases, and the LIME approach yielded explanations in line with marketing knowledge.
DOI: https://doi.org/10.5281/zenodo.20233483
