Online commerce currently provides better customer activity information, compared with traditional retail stores. For instance, measuring customers' interest on products in shelves is a complex task in physical environments. However, these scenarios may benefit from the Internet of Things (IoT) technologies to obtain context-aware information hard to obtain otherwise. For instance, in a real store, users may show their interest in a given product depending on the time interacted with it. We present a system designed to reliably detect user-object interactions in an RFID-enabled context-aware shelf scenario, with the goal to measure user activity based on the weighted Information Gain classifier (wIG), an empirical machine learning technique. The system is configured by means of thresholds determining the classification accuracy, and it is automatically adapted to different scenarios by means of an automated calibration method. Our proposed user-object interaction measurement method achieves performance above 80% in a real environment evaluation, indicating a high reliability. Our proposal could be used to feed user-centric privacy-preserving recommender systems in brick-and-mortar stores, or as aiding tool for visually impaired users.