Representation learning via Dual-Autoencoder for recommendation

Abstract

Recommendation has provoked vast amount of attention and research in recent decades. Most previous works employ matrix factorization techniques to learn the latent factors of users and items. And many subsequent works consider external information, e.g., social relationships of users and items’ attributions, to improve the recommendation performance under the matrix factorization framework. However, matrix factorization methods may not make full use of the limited information from rating or check-in matrices, and achieve unsatisfying results. Recently, deep learning has proven able to learn good representation in natural language processing, image classification, and so on. Along this line, we propose a new representation learning framework called Recommendation via Dual-Autoencoder (ReDa). In this framework, we simultaneously learn the new hidden representations of users and items using autoencoders, and minimize the deviations of training data by the learnt representations of users and items. Based on this framework, we develop a gradient descent method to learn hidden representations. Extensive experiments conducted on several real-world data sets demonstrate the effectiveness of our proposed method compared with state-of-the-art matrix factorization based methods.

Introduction

In order to tackle the information overload problem, recommender systems are proposed to help users to find objects of interest by utilizing the user–item interaction information and/or content information associated with users and items. Recommender systems have attracted much attention from multiple disciplines, and many techniques have been proposed to build recommender systems (Adomavicius and Tuzhilin, 2005, Bell, 2011). It is also widely used in many E-commerce companies, such as for product sale on Amazon and movie rentals from Netflix (Srebro & Jaakkola, 2003).

Traditional recommender systems normally only utilize user–item rating feedback information for recommendations. Moreover, the matrix factorization technique is widely used in recommender systems, which factorizes a user–item rating matrix into two low rank user-specific and item-specific matrices, and then utilizes the factorized matrices to make further predictions (Koren et al., 2009, Srebro and Jaakkola, 2003). In order to comprehensively utilize rich information in recommender system, there is a surge of hybrid recommendation, such as social recommendation (Ma et al., 2008, Ma et al., 2011), location based recommendation (Lian et al., 2014, Liu et al., 2013), and heterogeneous network based recommendation (Shi et al., 2015, Yu et al., 2014). Most of these methods are based on a matrix factorization framework, in which the latent factors of users and items are usually obtained by directly factorizing the user–item rating matrix and additional information are usually used as a regularization constraint. Although these methods pay much attention to exploit additional information, we wonder they might not make full use of the user–item rating information. In other words, we might be able to obtain better latent factors of users and items through extensively exploiting rating information.

On the other hand, deep learning has shown its power in learning latent feature representation in many domains, such as image/video processing (Alex et al., 2009) and text data (Socher, Huang, Pennington, Ng, & Manning, 2011). Can we use deep learning techniques to learn latent representations for recommendation? Some researchers have pursued this goal. For example, the latent factor of music is extracted from audio signals with a deep convolutional neural network (Oord, Dieleman, & Schrauwen, 2013) and the tabular data is modeled through the adaption of Restricted Boltzmann Machines (Salakhutdinov, Mnih, & Hinton, 2007). Recently, Wang, Wang, and Yeung (2014) designed a collaborative deep learning method to utilize item content information. In essence, these methods utilize the powerful representation learning of deep learning to analyze the additional information (e.g., audio signal and text context), not rating information, for better recommendations. They did not directly learn the latent factors of users and items with deep learning. Moreover, the additional information sometimes is not easy to acquire and very sparse.

To the best of our knowledge, there has been little effort focused on employing deep learning for recommendations only on user–item rating information. Motivated by the success of the latent feature representation of deep learning on image and text data, we design a novel Recommendation framework via Dual-Autoencoders (ReDa), which is illustrated in Fig. 1. In this figure, ReDa simultaneously learns the new hidden representations of users and items using autoencoders, and minimizes the deviations of training data by the learnt representations of users and items. Moreover, a gradient descent method is derived to learn the hidden representations. Experiments on four real-world data sets demonstrate the effectiveness of our proposed model.

The remainder of this paper is organized as follows. We introduce the notations and preliminary knowledge in Section  2, and then propose the representation learning framework based on autoencoders for recommendation in Section  3. Extensive experiments conducted on several data sets are shown in Section  4, followed by the related work in Section  5 and conclusions in Section  6.