Hierarchical Factorization Machines

• Applicant: Professor Dr. Ulrik Brandes, since 11/2013
• Subject Area: Image and Language Processing, Computer Graphics and Visualisation, Human Computer Interaction, Ubiquitous and Wearable Computing
• Term: from 2013 to 2017
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 229178042

Factorization models are one of the most successful approaches in machine learning and data mining for prediction problems over categorical variables of large domains as for example in recommender systems. Many different single-layer and hierarchical factorization models have been proposed for solving such prediction tasks. Each new model requires to derive a model equation and to develop a learning algorithm. Moreover, enhancements made for one factorization model are not directly applicable for the other models. Clearly, there is a lack of a generic model to subsume all these specialized factorization models.This project aims at bringing different factorization models together by developing a generic hierarchical approach. Therefore, (1) factorization machines (FM) which subsume most of the non-hierarchical factorization models are extended with Bayesian inference. (2) A multilevel hierarchical layout of several factorization machines including an inference method are investigated which should serve as a generic approach to model deep and complex prior structures. (3) The practical importance of HFMs is explored in the fields of recommender systems and social network analysis. Studying and developing such a generic model class will facilitate both future basic research as well as the applicability in practice.

• DFG Programme: Research Grants
• Ehemaliger Antragsteller: Professor Dr. Steffen Rendle, until 9/2013