From Sparse Ratings to Rich Representations: Multi-Source User Modeling for Cross-Domain Recommender Systems

Key Points

• The central aim is to address data sparsity issues in recommender systems by utilizing multi-source user information through a Doc2Vec-based framework.
• Developed a cross-domain recommendation framework using Doc2Vec to generate user-item embeddings.
• Combined user embeddings from source and target domains for common users.
• Utilized a multilayer perceptron to learn user-item interactions and predict ratings in the target domain.
• Conducted various experiments on real data from Yelp.com to evaluate model performance.
• The proposed approach showed superior predictive performance compared to baseline models using single domain information.
• Key factors affecting predictive performance included the number of common users and the number of reviews in the source domain.

Abstract