A Behavioral Economics-Driven AI Framework for Purchase Propensity Modeling in E-Commerce

Abstract

This paper presents a deep learning predictive control model that predicts purchasing intentions on e-commerce considering sentiment analysis, time-varying review dynamics, and the concept of behavioral economics. To model the actual consumer decisions, we developed a hybrid architecture of BERT, DeBerta and LSTNet, which is enhanced with risk aversion, expected utility and prospect theory. BERT is able to find semantic and emotional context of the user reviews, DeBERTa helps with understanding the context regarding the user reviews, and LSTNet can track the sentiment change over time, which represents how the users change preferences related to different products. The behavioral layer varies predictions based on an impression of risk, gain/loss framing and utility expectations, which provides a more real model of the choices in the purchase. The model is trained and evaluated using the Amazon Reviews data (2018), where the structured metadata price, brand and product attributes are added to the textual representation and processed together. As it can be seen in the results of the experimental work, the offered approach not only achieves competitive predictive results in terms of accuracy, F1-score, recall and AUC, but also delivers behaviorally valuable insights that could guide retailers to optimize pricing, recommendations along with inventory management. Integrating behavioral economics and deep learning means that the framework is new and relevant in enhancement of decision support system in e-commerce. The accomplished research revolves around user-centered predictive analytics, which is a multi-modal approach to online retail business and provides practical suggestions and enhanced congruence to actual consumer behavior.