In this thesis, we work on designing new recommender systems aligning with real scenarios. Traditional recommender systems have been witnessed its prosperity in different domains, e.g., e-commerce, movie and music online platforms, and social networks. Techniques proposed for accuracy-based recommendation algorithms perform comparatively well to prediction results precisely. However, with a wider use of recommender systems in healthcare, geo-location related domains where sensitive data should be carefully processed and used, employing merely accuracy-based metrics to measure how good a recommender system isfar from enough.

Recently fairness, diversity, interpretability and transparency of recommender systems have been brought up to evaluate a new generation of recommender systems. Fairness and diversity aim at providing users with a well-balanced result from recommender systems, reducing and avoiding bias effects and “filter bubbles” phenomenon. Interpretability and transparency aim at a more comprehensible system for users, which in turn will enhance the trust and ameliorate the acceptance of such systems with regards to sensitive data. Existing systems do not integrate all these aspects.

In this thesis, we address the diversity aspect of recommendation. We investigate what effects on diversity can be brought along by knowledge graph embedding techniques, how to achieve the best accuracy-diversity trade-off while considering not only historical user-item interactions but also auxiliary item information. We explore the use of Determinantal Point Processes (DPP) to achieve a better accuracy-diversity trade-off.

For evaluation purpose, we use real world datasets: MovieLens and Anime, and evaluate the effectiveness of proposed approaches from three perspectives: accuracy, diversity, accuracy-diversity trade-off.

Mots-clés : recommender system, diversity, accuracy-diversity trade-off, DPP