A Hybrid Approach for Improving Recommendation Systems

Abstract

Recommender systems have become ubiquitous on the internet. On most e-commerce and digital service platforms, users frequently encounter suggestions such as “customers who liked this product also liked that one.” The primary goal of these systems is to personalize the browsing experience, optimize the conversion rate (visitor → customer), and facilitate the retrieval of relevant information from massive volumes of data. However, despite their effectiveness, recommender systems face several challenges: (1) the lack of data truly adapted to users’ specific needs, (2) user disorientation during the search process, leading to inefficiency, and (3) difficulties in leveraging the richness of multimodal data collected from diverse sources such as social networks, reviews, interaction histories, and contextual information. Recent research has attempted to address these issues by integrating machine learning and deep learning techniques, as well as multi-criteria decision-making methods, to better capture user preferences, context, and external factors.

To overcome these limitations, we propose in this work two complementary recommender systems. The first, the Smart Food and Restaurant Advisor (SFRA), employs Multi-Criteria Decision-Making (MCDM) methods to support healthier and more personalized food choices. Unlike traditional approaches that neglect nutritional and lifestyle dimensions, SFRA integrates user profiles enriched with dietary needs, preferences, and geographic data to deliver context-aware recommendations. Our system evaluates and ranks food and restaurant options by considering nutrition, ingredients, and contextual relevance, enabling users to make informed choices that balance taste with health goals. The second system is an innovative multimodal recommendation approach that exploits heterogeneous data sources (user preferences, spatio-temporal context and emotional) and leverages Graph Neural Networks (GNNs) to model complex relationships between users, items, and contextual factors. Furthermore, we introduce a multi-list recommendation mechanism, designed to provide multiple personalized suggestion lists based on different criteria (relevance, diversity, context), thereby enhancing personalization while reducing user disorientation.