DNR: A Unified Framework of List Ranking With Neural Networks for Recommendation

In recent years, list ranking learning has attracted further attention in the field of recommendation. Most of the existing list ranking methods use traditional or deep learning methods to fit user-item interactions and output a personalized ranking list. These solutions mainly focus on a certain aspect of linear and nonlinear interactions when learning latent structures of users and items. In fact, linear interactions and nonlinear interactions coexist and play complementary roles in a recommendation system. Once any type of interaction is ignored, it will result in the loss of linear or nonlinear features, which will further affect the overall performance of the model. To address this problem, we propose a general framework, DNR, short for Deep Neural Rank. To jointly learn linear and nonlinear features, it uses both traditional and deep learning methods to fit user-item interactions. This is a flexible architecture that can not only be extended to the integration of various linear models and nonlinear models but also be simplified for pairwise learning to rank. This paper focuses on the feasibility of integrating matrix factorization (MF) and multi-layer perceptron (MLP) into the framework as linear and nonlinear methods. Eventually, we conduct extensive experiments on three data sets (Movielens100K, Movielens1M, and Yahoo Movies). The results show that our proposed DNR framework achieves a significant improvement compared to existing methods.

Automated summary

Most existing list ranking methods focus on learning either linear or nonlinear interactions, while ignoring the coexistence and complementary roles of both. To address this issue, the proposed DNR framework combines traditional and deep learning methods to jointly learn linear and nonlinear features for improved recommendation system performance.