Shared-view and specific-view information extraction for recommendation
Introduction
With the explosive growth of information, recommendation systems are becoming more and more important in alleviating information overload, and have been widely used in many websites and applications. Recommendation systems help customers make decisions by displaying candidate products that they may be interested in, according to their preferences and past purchasing behaviors. Many important methods widely used in recommendation systems are collaborative filtering (CF) based techniques (Berg et al., 2017, Panagiotakis, Papadakis, and Fragopoulou, 2021), which use historical records (such as ratings and clicks) to properly model user preferences and product features. These methods assume that users who have similar choices in the past will be interested in similar products in the future. Many CF-based methods decompose the rating matrix into two latent feature matrices of users and items according to Matrix Factorization (MF), and then predict the missing rating of the user to the item with the inner product of user and item latent feature vectors. Those methods can find some common factors, which may be the potential reasons why users like the item. For example, when recommending a book to users, these factors can be the author or genre of the book that users express concern about. Besides finding these hidden factors, MF technology can also show how each item meets each factor and how important they are to each user. But the performance of those CF-based recommendation systems decreases significantly, when there is little user–item interaction information in the recommendation system. In this case, the neighbor-based methods are introduced to tackle the problem of data sparsity, which use various relationships between entities, such as items or users, to make the recommendation more accurate.
In social media sites and e-commerce systems, some users further express their preferences by writing reviews after they have consumed the items. These reviews contain rich information, which can be exploited to alleviate the problem of data sparsity. Some studies (Dong et al., 2020, Jakob et al., 2009, Zheng et al., 2017) show that the methods taking reviews into consideration usually outperform the CF-based methods only considering interaction records. The user and item representations of these review-based methods are extracted from reviews. For example, Deep Cooperative Neural Networks (DeepCoNN) ( Zheng et al., 2017) obtains the embeddings of user preferences and item properties from all reviews of this user and item by using two parallel neural networks. However, some reviews contain noise and some contain too few words to extract effective information. The performance of those recommendation systems will be seriously damaged when these useless reviews are introduced into the model.
Ratings and reviews are two different channels of information to represent the preferences of users to items. For example, in movie recommendation, the preferences of users can often be represented by ratings and comments to movies. The rating may reflect how much the user is interested in one movie. While the reviews will explain the reasons from several aspects, such as the performance of actors, the style of the movie, and the plot. In this paper, we seek to seamlessly unify ratings and reviews to make recommendation.
There are two examples in Fig. 1. The first user gave the highest rating, and the reviews also showed that he was satisfied with the CD to some extent. Since reviews can express the opinions of the user from several aspects, we can find the user also pointed out some drawbacks of the CD even though he gave the highest score. The second user gave a very low rating, and he talked about the shortcomings of the disc in his reviews. But it can be seen from the reviews that he did not completely deny the disc. From Fig. 1, we can see that most ratings and the reviews are similar with each other from the emotion view, and may complement each other. We can extract shared information from the interaction data and reviews. However, ratings are numerical information while reviews are textual information, and they reflect the preference of the user from a coarse-grained level and a fine-grained level, respectively. That is to say, each of them contains their own specific contribution to the final rating prediction, and the heterogeneity between them cannot be ignored. The main challenge is how to integrate the two types of heterogeneities in an efficient and general way.
In fact, some works have combined reviews and the interaction data to extract more effective representations of users and items, and have proved that incorporating both of them into recommendation systems can further improve the recommendation performance. Wu, Quan, Li, Wang, and Zheng (2019) propose a context-aware user–item representation learning model (CARL), which integrates interaction information and text reviews to make rating prediction. However, CARL encounters a problem. When trying to combine interaction data and reviews information, CARL firstly maps each kind of information into its own subspace, and then concatenates the representation of each subspace to predict ratings. This process is always carried out in an independent way, and the complementary and inherent relation between ratings and reviews has been ignored. Hybrid Recommendation model to learn Deep Representation (HRDR) (Liu, et al., 2020) takes their relationship into account, and the representation learned from ratings is used as an attention query vector to select shared information from reviews. But the heterogeneity between reviews and ratings makes HRDR unable to obtain some useful reviews via rating-based selection, such as the sentences marked in blue in the first example and those marked in red in the second example in Fig. 1, and those reviews also make specific contributions to enhance rating prediction.
In order to address the above issues, we propose a Dual-View Learning model with Shared-view and Specific-view Information extraction for Recommendation (SSIR), which integrates the information from reviews and interaction matrix to predict ratings. The model SSIR has two key components, they are shared subspace information extraction and specific-view exploitation parts, separately. From the perspective of shared-view, SSIR jointly minimizes the loss of confusion adversarial and rating prediction loss to extract the shared information from reviews and user–item interaction matrix. The distributions of shared features extracted from ratings and reviews will become as similar as possible when we minimize the loss of confusion adversarial. For the specific-view part, the useful specific information is hard to define. However, we can extract it by eliminating the shared information from original information. This can be achieved by applying orthogonal constraints on shared-view and specific-view modules, to guarantee there is no correlation between these two modules. Then, we concatenate the shared-view and the specific-view representations to gain the final embedding for the user and item. And the final embedding is fed into a Factorization Machine (FM) for rating prediction. We have carried out comprehensive experiments on eight real-world datasets. The experimental results show that SSIR is superior to several state-of-the-art algorithms. The results also validate that the Shared-view and Specific-view Information can complement each other, and fusion of them will make a better prediction performance. The major contributions of our work are summarized as follows:
- •
We put forward a Dual-View recommendation model (SSIR), which extracts information from the shared view and their specific views, to fully mine the information in the reviews and the interaction matrix.
- •
In the shared view module, we utilize the confusion adversarial loss to extract the shared features from different kinds of information, such as reviews and the interaction data. As far as we know, we are the first to apply this idea in the recommendation scenario. We also enforce orthogonal constraints on shared and specific subspaces to extract the information from specific views. Finally, shared-view and specific-view information are synergized to learn the final embeddings of users or items for rating prediction.
- •
We empirically conduct extensive experiments on eight real-world datasets. Experimental results demonstrate that our model SSIR outperforms several state-of-the-art alternatives. Moreover, some ablation studies verify the effectiveness of our model for recommendation and the utility of our model in alleviating data sparsity.
Section snippets
Related work
In this section, we briefly review the related work from two different perspectives. The first category discusses the relevant literatures on collaborative filtering. The second category covers advanced studies about reviews-based recommendation.
Preliminaries
We first give the formulation of the problem in Section 3.1. Then we introduce how to initialize ID embeddings of users and items in Section 3.2. In Section 3.3, we shortly recapitulate the CNN text processor.
The proposed model
Fig. 3 illustrates the overall framework of SSIR, which consists of three components: (1) shared-view information extraction module, which gets the shared features from reviews and interaction data; (2) specific-view information extraction module, which captures the discriminative features from reviews and interaction data; and (3) rating prediction module, where the final comprehensive representations are used to estimate the a user’s rating to an item. The framework of SSIR is symmetric. The
Experiments
In this section, we use eight real-world datasets from two different websites to conduct extensive experiments for performance evaluation. Then, we analyze the influence of different parameter settings of SSIR on the experimental results and the contribution of different components to the performance of SSIR.
Conclusion and future work
In this paper, we propose a Dual-view recommendation model SSIR, which can extract effective representations from ratings and reviews. The key of our model is to extract shared view and specific-view information from ratings and text reviews. Specifically, we jointly minimize the loss of confusion adversarial and rating prediction loss to extract the shared information in the shared-view module. Then, we apply orthogonal constraints on the specific-view module and shared-view module to extract
CRediT authorship contribution statement
Huiting Liu: Conceptualization, Methodology, Writing – review & editing. Jindou Zhao: Software, Validation, Formal analysis, Data curation, Writing – original draft, Visualization. Peipei Li: Investigation. Peng Zhao: Resources. Xindong Wu: Supervision.
Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
Acknowledgments
This research has been supported by the National Key Research and Development Program of China No. 2016YFB1000901, the National Natural Science Foundation of China Nos. 61202227 and 61602004, Natural Science Foundation of Anhui Province, No. 2008085MF219 and Provincial Natural Science Foundation of Anhui Higher Education Institution of China , No. KJ2018A0013.
References (40)
- et al.
Hybrid neural recommendation with joint deep representation learning of ratings and reviews
Neurocomputing
(2020) - et al.
Improving recommender systems via a Dual Training Error based Correction approach
Expert Systems with Applications
(2021) - et al.
Joint representation learning with ratings and reviews for recommendation
Neurocomputing
(2021) - et al.
Multi-criteria recommender system based on social relationships and criteria preferences
Expert Systems with Applications
(2021) - et al.
Graph convolutional matrix completion
(2017) - et al.
TransNets: Learning to transform for recommendation
- et al.
Sequential recommendation with user memory networks
- et al.
Try this instead: Personalized and interpretable substitute recommendation
- et al.
Attentive collaborative filtering: Multimedia recommendation with item- and component-level attention
- et al.
Neural attentional rating regression with review-level explanations
Jointly modeling aspects, ratings and sentiments for movie recommendation (JMARS)
Attentive aspect modeling for review-aware recommendation
ACM Transactions on Information Systems (TOIS)
An effective distributed predictive model with matrix factorization and random forest for big data recommendation systems
Expert Systems with Application
Adversarial personalized ranking for recommendation
NAIS: Neural attentive item similarity model for recommendation
IEEE Transactions on Knowledge and Data Engineering
Neural collaborative filtering
Beyond the stars: exploiting free-text user reviews to improve the accuracy of movie recommendations
Convolutional matrix factorization for document context-aware recommendation
Grouplens: applying collaborative filtering to usenet news
Communications of the ACM
Cited by (2)
Enhancing review-based user representation on learned social graph for recommendation
2023, Knowledge-Based SystemsHow do context-aware artificial intelligence algorithms used in fitness recommender systems? A literature review and research agenda
2022, International Journal of Information Management Data InsightsCitation Excerpt :By proposing a probabilistic MF (Pujahari & Sisodia, 2020), user and side information about items are integrated into the model. In another study, Liu et al. (2021) propose a shared-view and specific-view information extraction model for recommendation (SSIR) by forming an interaction matrix of both numerical and textual views to predict ratings. Wang et al. (2019) suggest “knowledge graph convolutional networks (KGCN)” that exploit the connectedness of attributes of a neighborhood through a knowledge graph.