2021/04/26 by Martin Milenkoski, Diego Antognini, Milenkoski, Martin +3
Computer Science · #Recommender Systems and Techniques #Topic Modeling #Multimodal Machine Learning Applications
paper · pdf · doi:10.48550/arxiv.2104.12822
In this paper, we describe a method to tackle data sparsity and create\nrecommendations in domains with limited knowledge about user preferences. We\nexpand the variational autoencoder collaborative filtering from a single-domain\nto a multi-domain setting. The intuition is that user-item interactions in a\nsource domain can augment the recommendation quality in a target domain. The\nintuition can be taken to its extreme, where, in a cross-domain setup, the user\nhistory in a source domain is enough to generate high-quality recommendations\nin a target one. We thus create a Product-of-Experts (POE) architecture for\nrecommendations that jointly models user-item interactions across multiple\ndomains. The method is resilient to missing data for one or more of the\ndomains, which is a situation often found in real life. We present results on\ntwo widely-used datasets - Amazon and Yelp, which support the claim that\nholistic user preference knowledge leads to better recommendations.\nSurprisingly, we find that in some cases, a POE recommender that does not\naccess the target domain user representation can surpass a strong VAE\nrecommender baseline trained on the target domain.\n