Joint Deep Recommendation Model Exploiting Reviews and Metadata Information

Abstract

User-generated product reviews contain a lot of valuable information including users’ opinions on products and product features that is not fully exploited by the current recommendation models. Similarly, the metadata information related to the products, about the reviews and about the users who have written the reviews has rarely been exploited for recommender systems. These heterogeneous information sources have the potential to alleviate the cold start and sparsity problems and improve the quality of recommendations. In this paper, we present a joint deep recommendation model (JDRM) that consists of two parallel neural networks, learning lower-order feature interactions of users and items separately and higher-order feature interactions jointly using a shared last layer. Each of the networks is further composed of two sub-networks. One of the sub-networks focus on exploiting product reviews (of user/item) and the other sub-network learns user preferences/items properties leveraging metadata information along with the ratings. The learned latent features in each network are concatenated, thus producing the user and item latent feature vectors. We combine the two networks by introducing a shared layer on the top, which is a dense fully connected layer used to learn higher level latent features obtained from the two networks and produces final ratings. Extensive experiments on real-world datasets demonstrate that JDRM significantly outperforms state of the art recommendation models.

Introduction

During the last decade, the number of products, the variety and the services provided by the companies have increased exponentially. Companies produce a large number of products along with enormous information to address customers’ needs. Although this provides more options to customers, but due to increasing information load it makes rigorous for them to render large amount of information provided in order to acquire their desired product needs. Recommender systems help customers to cope this problem of information overload by recommending the products that are probably of their interests based on their needs, preferences and past buying behavior [1]. Now a days, various applications like Amazon, Netflix, and Twitter etc. use recommender systems to help people in their daily life.

In the current era of smart devices and web 2.0, users write textual reviews on many e-commerce systems to express their personal opinions towards the purchased items. They also provide rating score for the products along with the reviews indicating their preferences. The e-commerce systems, like Yelp and Amazon, also maintain the metadata information about the users, products and about reviews written by the users. The textual reviews and metadata information may contain the information related to properties of items, users’ behavior and also the preference of each individual user. Before the exploitation of deep learning techniques in this very area, collaborative filtering (CF) with matrix factorization (MF) [2], [3] was considered to be state-of-the-art technique. This technique is built on the basic idea that people having similar choices in the past are likely to have similar interests in the future [39]. Since, dot product is used in MF which has strong constraint that the latent dimensions are independent of each other i.e. each dimension in latent user vector would only interact with the corresponding dimension in the latent item vector. Thus, the MF based CF technique is incapable of learning complex user-item rating behaviors through higher order feature interactions. Although this technique had shown good performance for recommender systems, however lacking in addressing data sparsity and cold start problems as it uses only explicit ratings given with the reviews. One of the techniques used to cope with the issue of lack of data is exploiting the information from the text in reviews. Recently, some researches [4], [5] have demonstrated that the performance of recommender systems can be improved using review text, particularly in case of cold start problems. Some other works [6], [7] utilized both purchased item review data and user-item interactions to improve recommendations, but jointly using product reviews and metadata has not yet been exploited.

In this paper, we propose a Convolutional Neural Network based model, called Joint Deep Recommendation Model (JDRM), to model user preferences and item properties by jointly using the text from reviews and the associated metadata for rating prediction problems. JDRM learns the latent feature vectors for users and items jointly from the two information sources by using two parallel neural networks such that the performance of the proposed model is maximized. One of the networks exploits the reviews, and the other network uses metadata to model user preferences and item properties. This architecture is inline with [8] where different features exploited jointly by a deep neural network in a query based interactive recommendation system. The learned latent features from product reviews and metadata are then feed (after merging) to the fully connected layer introduced on top of the two networks. The most influential features (higher-order) form the latent (specialized) features are captured in this layer, which are later passed through a regression layer to estimate the rating. Our work is novel in the sense that the proposed model, to the best of our knowledge, is the first one to capture both user preferences and item properties from different information sources in a unified neural way, and we are exploring for the first time the metadata information (which is now commonly available with many e-commerce systems like Yelp and Amazon) along with the review text. A substantial advantage of JDRM is that it performs modeling in an integration manner as compared to numerous other approaches [4], [5], [7] that benefit from reviews and contents of users/items. As most of the similar approaches in literature learn latent features from different information sources (product reviews, item descriptions, user-item ratings) independently, thus there are lesser chances of learning good features that can be useful for rating prediction. We believe this work may be impactful in two scenarios for research community and e-commerce systems: (i) Many different kind of metadata are now managed along with the reviews by different e-commerce systems like Yelp and Amazon. Our work validates the usefulness and impact of metadata when used together with the review text from product reviews, and its exploration urges the researcher community to explore further. (ii) The kind of metadata we used (from Yelp) will encourage other recommender / e-commerce systems to maintain similar or even richer information so that their systems can provide better recommendations after applying artificial intelligence.

The experimental results demonstrate that JDRM outperforms all the baseline methods in prediction accuracy on the evaluated real-world datasets of Yelp. Also, the proposed model exhibits good performance for cold start issues and alleviates the sparsity problem by leveraging product reviews and the associated metadata. In summary, the main contributions made in this paper are listed as follows:

• We propose a novel joint deep learning model, named JDRM that models user preferences and item properties using product reviews and metadata in an integration manner. The fusion/joint layer on top of the two parallel networks connects them in a way that the most influential features from the two latent representations can be drawn for the final rating predictions. JDRM is the first attempt that exploits products reviews and metadata using neural networks for modeling user preferences and item properties.

• We demonstrate the use of metadata information empirically. From the results we can see that, on all the datasets, the performance of the proposed model is much better than the performances of the two networks trained separately. It justifies the unified use of the two information sources in the proposed way.

• The experimental results demonstrate that JDRM achieves superior performance in terms of prediction accuracy on publicly available benchmark datasets, and outperforms a variety of state-of-the-art baseline methods [9], [10] (see Section 4).

The rest of the paper is organized as follows. Section 2 provides a brief literature study related to this work. JDRM is described in detail in Section 3. Experiments are presented in section 4 to examine JDRM and validate its effectiveness in comparison with the baseline methods for recommender systems. Finally, conclusions are presented in Section 5.