As e-commerce shifts from an information economy to an attention economy, merchants increasingly compete to attract and retain consumer attention in crowded digital marketplaces. Generative AI, particularly large language models, is transforming how merchants interact with customers by enabling personalized, human-like, and context-aware engagement. This transition moves e-commerce away from static, search-based systems toward dynamic, agent-based interactions. While prior research has examined customer attention through various methods, such as attention allocation models, eye-tracking, cognitive load theory, and behavioral economics, these approaches struggle to integrate multimodal data, contextual and emotional factors, and social dynamics. Traditional predictive and decision-making models also lack natural language understanding and generation capabilities, which limits their effectiveness in capturing customer intent and supporting adaptive, attention-driven strategies.
This project aims to develop AI-based agents for attention-driven e-commerce applications that support real-time engagement, adaptive responses, and autonomous decision-making. In the first phase, generative AI will be used to uncover new patterns in customer attention and behavior from multimodal data, followed by the development of few-shot (limited examples) and zero-shot (no prior examples) predictive models for merchants. Building on these insights, the project will design adaptive, attention-driven product selection and sales strategies, combining large models for sentiment analysis, intent recognition, product positioning, and content generation with smaller, more efficient models, such as dynamic programming and machine learning, to optimize keyword selection, targeted advertising, and personalized sales across organic and paid channels.
In the second phase, the project will develop task-oriented AI interactive optimizers and agent-based decision-making frameworks tailored to e-commerce operations. These optimizers will incorporate both natural language and numerical feedback to assist merchants with specific tasks, including product assortment optimization, adaptive matching and rationing decisions. Agent-based modeling and mixture-of-experts architectures will be explored to support dynamic conversations and complex environment simulations. The project will further introduce a disciplined supervised fine-tuning framework for diffusion models using continuous-time reinforcement learning, formulated as a stochastic control problem that aligns model outputs with merchant objectives, ultimately enabling robust, efficient, and scalable AI-driven decision support for e-commerce merchants.
