A Study on the Multi-Objective Recommender Systems Based on Deep Learning(2/3)

This project plans to leverage on the deep learning technology to build a multi-objective recommendation model, which relaxes the constraints of the matrix factorization-based recom- mendation models. The matrix factorization-based recommendation models at least have the following problems.1. MF-based models assume that the weight of each latent factor is the same, which may not always be a reasonable assumption.2. MF-based models assume that each latent factor is independent of other factors, which may not always be a reasonable assumption.3. MF-based models require the number of latent factors of a user equaling the number of latent factors of an item. However, this may not be a good encoding mechanism.4. We probably over-simplify the interaction between a user and an item if only applying the inner-product operation. Particularly, we should consider higher-order interactions, such as the kernel method or the deep learning related operations.This research attempts to design the recommendation model based on the deep learning architecture to relax these restrictions. For the first issue, we have already developed a prototype such that each latent factor has distinct weight. Initial experimental results show that the new method performs better than the other MF-based approaches on the test dataset in terms of the root-mean-squared-error (RMSE).In addition to recommender systems, MF-based approaches can be applied to a wide range of applications, such as link prediction, imputation of the missing values, image compression, etc. This study may relax the constraints of the traditional MF-based approaches, so the results can also be applied to the above domains. Since the new techniques are more general than the MF-based approaches, we may get better results on the above applications.