Leveraging the Subtle: Hidden Factors in Recommender Systems

Abstract

Recommender systems are pivotal in various domains, aiding users in their decision-making. However, current systems often overlook subtle factors that significantly impact user preferences and choices. This work aims to bridge this gap by exploring the conceptof implicit item characteristics -- latent features that influence user decision-making in addition to explicit content. The investigation is divided into three key research areas. Firstly, we explore how to systematically identify and expose implicit item characteristics to enhance recommender systems in two key domains: tourism and news. Using advanced analytics such as cluster analysis and multiple linear regression, we map tourist destinations to the established Seven-Factor Model in tourism. In the news, we employ natural language processing techniques to reveal hidden features essential for tailoring recommendations. Secondly, we introduce a novel system called PicTouRe to elicit tourists' implicit preferences through pictures. Leveraging convolutional neural networks, we translate visual preferences into a Seven-Factor profile for each user, simplifying decision-making and capturing both immediate touristic desires and enduring personality traits. Lastly, we enhance news recommender systems by leveraging sentiment and emotions of news articles. Two models, RobustSentiRec and EmoRec, were developed to capture these implicit characteristics, aligning recommendations more closely with user preferences but also raising ethical concerns around potential sentiment and emotional echo chambers. Our findings offer a robust framework for more nuanced, user-sensitive recommendations, opening new avenues for future research and applications in recommender systems.