Learning Hybrid Recommender Models for Heterogeneous Semantic Data

AutorAndreas Lommatzsch, Benjamin Kille, Sahin Albayrak
QuelleACM SAC'13, March 18-22, 2013, Coimbra, Portugal 
LinksBibTeX   |   Uni-Bibliothek 

Recommender algorithms are key technologies supporting users to deal with information overload. Recommender Systems (RS) aim at identifying relevant items a user is unaware of. The relevance of items depends on various criteria such as item properties, the relationship to other items, user preferences, and contexts. The aggregation of different criteria is the key in computing high-quality recommendations. We propose a novel approach for automatically learning parameters for recommendations based on semantic datasets (graphs). We outline how scaling models and noise reduction models allow us to consider individual dataset properties. We evaluated the proposed methods on a semantic movie data set. The evaluation shows that each semantic relationship set requires a separate recommender model. Combining such recommender models yields much higher precision. We show the recommender ensembles outperform recommenders based on aggregated semantic graphs (block matrix recommender).