Intelligent Taste Prediction Engine Using Scroll, Hover & View-Time Patterns in E-Commerce

Abstract: E-commerce platforms usually recommend products by looking at a user’s past activities, such as the items they clicked on, rated, or purchased earlier. While this works well for regular users, it does not clearly understand what a user wants during live browsing. This problem is more serious for first-time visitors and users who are not logged in, often resulting in general recommendations, low user engagement, and higher cart abandonment. Although advanced models like deep learning, graph-based methods, and transformers improve accuracy, they need large amounts of historical data, heavy computation, and offline training. Because of this, they are not practical for medium-scale e-commerce platforms. To solve these issues, this paper introduces a lightweight, intent-aware recommendation system that works in real time. The system observes small but meaningful user actions such as how far a user scrolls, how long they hover over products, how much time they spend viewing items, and how often they switch between products. Using these signals, the system understands the user’s intent during the same browsing session and updates recommendations instantly. It does not depend on personal or stored user data, which helps protect user privacy. Experimental results show that the proposed system achieves around 88–90% accuracy, reaching nearly 90% of the performance of complex existing models, while using much less computing power. This makes the system efficient, privacy-friendly, and suitable for real-time use in small and medium-scale e-commerce applications.

Keywords: E-commerce recommendation systems, session-based recommendation, real-time personalization, micro-interaction analysis, user intent prediction, lightweight machine learning, cold-start problem, privacy-preserving recommendation, small and medium enterprises (SMEs).