A Privacy-Preserving Hypercube Framework for Interactive Recommendations

Interactive decision-support systems increasingly rely on cloud-based recommender architectures, which introduce latency, energy consumption, and privacy concerns that may negatively affect user experience and trust. In this paper, we present a content-based, on-device recommendation framework in which all inference is performed locally, without server-side computation or data sharing. The approach is based on a lightweight hypercube representation of items, enabling efficient and interpretable personalization. The system is designed to operate across multiple domains, including movies and restaurants, and supports extensions that incorporate side information, such as features extracted from textual reviews. We report initial qualitative findings from a user study in the movie domain, focusing on users’ perceptions of usability, personalization, and privacy.