AI-Driven Personalized Shopping Recommendation Framework Using Small Language Models and User Behavior Analytics

Abstract

In today's e-commerce landscape, personalized recommendation is a critical tool for enhancing customer engagement and decision-making. The traditional recommendation methods, however, have the disadvantages of weak user interaction, cold start issue, insufficient contextual information and high computing demand of large language models. This study introduces the concept of an AI-Driven Personalized Shopping Recommendation Framework, combining user behavior analytics and Small Language Models (SLMs) to effectively recommend products based on user context. The framework examines visitors' purchase frequency, browsing habits, search history, dwell time, category preferences, brand affinity, and price sensitivity to build dynamic user profiles. SLMs are used for contextual preference summarization, semantic product understanding and personalized recommendation generation. A hybrid ranking method is used which integrates behavioral, semantic similarity, SLM preference scores and product popularity, and an explainability layer is created to generate recommendation reasons and confidence scores. The experimental results prove that the framework outperforms conventional recommenders and deep learning recommenders in terms of the recommendation quality, as well as the inference latency and computational requirements, thus making it a potential solution for a real-world e-commerce application that needs to be scalable.