Deep generative models for collaborative filtering with multi-source and high dimensional data

Abstract

This thesis examines hybrid collaborative filtering, which is a class of the collaborative filtering methods with multiple high dimensional auxiliary information. Regarding methodology, we consider structured probabilistic models and deep generative models. In the first chapter, we look at structured probabilistic models for phenotyping with electronic health records in the field of medical informatics. In the second chapter, we explore the modeling aspects of deep generative models for hybrid collaborative filtering. Finally, the third chapter deals with a hybrid collaborative filtering model considering domain characteristics. We propose and compare structured probabilistic models and deep generative models for hybrid collaborative filtering. In particular, we cover various modeling aspects in terms of deep generative models and compare them empirically. These results can be useful guidelines for future research and are expected to contribute to the expansion of the relevant fields. We also propose a hybrid recurrent collaborative filtering model for sequential diagnosis prediction that considers the medical domain characteristics. This effort can contribute to overcoming the limitations of static collaborative filtering methods.