Multi-Channel and Fusion Encoding Strategy Based Auto Encoder Model for Video Recommendation

Wenjie Yan,Jing Liu,Zhi Li,Dong Wang,Liang Ma

doi:10.1109/access.2019.2925653

Wenjie Yan, Jing Liu + Show 3 more

Open Access

https://doi.org/10.1109/access.2019.2925653

Copy DOI

Abstract

Recommendation techniques are widely used in many areas to deal with the information overload problem. However, the recommendation theory has suffered from the sparseness problem, which decreases the precision of the recommendation algorithms heavily. Deep learning theory has proven to be a very efficient tool to mine the latent information of data. In this paper, a novel scalable multi-channel and fusion encoding strategy-based auto encoder (MCFE-AE) model is introduced to make recommendations by deeply mining the latent features of users and video items of the data. The detail of the proposed algorithm is summarized as follows. First, the rating data that represent the users’ preference are sent to the input port of the proposed MCFE-AE model as raw input data. Second, the latent features of users and items have been deeply mined by the multi-channel and fusion encoding process of the proposed MCFE-AE model. Third, the final rating prediction result has been obtained by the decoding process of the proposed MCFE-AE model. The extensive experiments have shown the benefits of the proposed algorithm on the measure of mean absolute error (MAE) and root mean square error (RMSE) compared with the state-of-the-art algorithms. Besides, the number of channels of the MCFE-AE, the $L_{1}$ and $L_{2}$ regularization method, the learning rate, and the important regularization parameter $\lambda $ have been studied thoroughly in this paper.

Highlights

With the development of internet and big data technique, data on the internet is rapidly increasing
Inspired by the local sensing principle of convolutional neural networks, we propose the multi-channel and mean fusion encoding strategy based on the auto encoder theory
The extensive experiments are conducted to verify the superiority of the proposed algorithm on root mean square error (RMSE) and mean absolute error (MAE) measure compared with the start-of-the-art algorithms, such as UserAverage, ItemAverage, RMB-DNN, topK-UBCF, topK-IBCF, NMF, PRA, SlopOne, NRR, PMF, Bayesian Probabilistic Matrix Factorization (BPMF), DAE-CTN, ReDa, BiasedSVD, NNMF, FML and I-AutoRec algorithm

Summary

Introduction

With the development of internet and big data technique, data on the internet is rapidly increasing. The phenomenon of massive data exists in all kinds of the social life, such as E-commerce, entertainment, medical treatment, education and so on. Recommender systems(RSs) have played an utmost role in internet era to solve the information overload problem [1]. RSs make predictions and recommendations of items to users by using different sources of information, such as users’ behavior, demographic features, social information, the information from internet of things(e.g., GPS locations, RFID) and so on [2]. The RSs can be roughly categorized into two categories, which are content-based and collaborative filtering(CF) [3]. Content-based RSs use the contents of items to analyze the similarities among them. User interests profile is established after analyzing sufficient number of the items

Objectives

Methods

Findings

Conclusion