[2Q5-J-2] Machine learning: advanced models

[2Q5-J-2-02] Constructing a Recommender System using Variational Neural Networks

〇Tomoki Ohtsuki¹, Shinsuke Sugaya¹, Ohnishi Yuki^1,2 (1. BizReach, Inc., 2. Purdue University)

Keywords:Variational Inference, Neural Network, Combinatorial Optimization

To build a recommender system, it is essential to infer the underlying latent features of users and items from the action logs (i.e., collaborative filtering), simultaneously taking their known attributes into consideration. For this purpose, numerous models have been proposed to naturally interpolate between these two extremes, especially in the context of extending the matrix factorization algorithm. In this work, based on the recent advances in deep learning technology and high-level probabilistic programming software, we propose a model which is extremely simple and easy to implement, yet flexible and robust against over-fitting. We also deploy the proposed model for a real service by combining its prediction with a combinatorial optimization program, and see a significant improvement in certain indicators.