Elsevier

Neurocomputing

Volume 442, 28 June 2021, Pages 307-316
Neurocomputing

Hybrid collaborative recommendation of co-embedded item attributes and graph features

https://doi.org/10.1016/j.neucom.2021.01.129Get rights and content

Abstract

In recent decades, personalized recommendation systems have attracted much attention from multiple disciplines for recommending interested products and services to users. Recommendation accentuates both the importance of feature learning tasks and the challenges posed by the sparsity of rating matrix. A common method for addressing the sparsity problem is to extend the feature space by the attributes of users and/or items. However, there are two main drawbacks in most existing recommendation methods. The first is the high computational cost of most existing recommendation models when using additional information from users and/or items to expand the feature space. The second problem is that it is difficult to obtain user additional information due to the high cost of acquiring tag knowledge and the increase in user privacy awareness. In this paper, we propose a novel and simple model to address the abovementioned issues, which employs a semi-autoencoder to co-embed the attributes and the graph features of the items for rating prediction (short for Item-Agrec). More specifially, a semi-autoencoder is introduced to learn the hidden nonlinear features of items for achieving a low computational cost, and thus the proposed Item-Agrec model can flexibly use side information from different sources. Meanwhile, in the case that it is not easy to obtain the user’s additional information, we take the item’s graph features and attributes into consideration for improving the accuracy of recommendation. Experiments on several real-world datasets demonstrate the effectiveness of the proposed Item-Agrec compared with state-of-the-art attribute-aware and content-aware methods.

Introduction

With the explosive growth of online information on the Internet, recommendation systems have played an indispensable role in overcoming the problem of information overload and have been widely applied in various fields. Recommendation system is an effective information filtering tool, which can find more interesting products and services from a large number of candidate items for users. Recommendation systems are generally divided into three types: content-based methods, collaborative filtering (CF) methods, and hybrid methods [1]. Content-based methods recommend items similar to the items he liked in the past for the user. The CF methods discover the user’s preference by mining the user’s historical behaviors, and recommend products or services that the user may be interested in. Hybrid methods are the combination of multiple recommended methods.

CF-based methods have achieved very competitive performance by virtue of its effectiveness and efficiency advantages. Most of the existing CF-based methods are based on matrix factorization (MF). However, matrix factorization still suffers from cold start and data sparsity. Some researchers have proposed latent factor analysis-based (LFA) models to solve the problem of sparse matrices [2], [3], [4]. Most existing LFA models are based on MF, and have defects in nonlinear features learning [5]. Recently, deep learning methods have shown the power on learning latent feature representation [6]. Especially, autoencoder-based methods are widely popular in recommendation systems for the superiority of no label requirement and fast convergence speed. The autoencoder-based AutoRec [7] model directly uses historical ratings to reconstruct the rating matrix, and obtains effective recommendation results. Wu et al. proposed the CDAE model, which uses the collaborative denoising autoencoder to learn the hidden features of users and items for top-N recommendations [8]. Autoencoders can effectively learn the hidden nonlinear features of users and items in personalized recommendation.

Most models that introduce additional information from users and/or items to solve the problem of data sparsity have achieved excellent performance [9], [10]. Many models cannot directly utilize any additional external features and require further modifications. This often leads to overly complex hybrid models, which may not be extended to simple scenarios where external information is scarce [9]. Due to the limitation that the dimensionality of the input layer of the autoencoder and the dimensionality of the output layer are consistent, the autoencoder-based model also faces this problem. To tackle this challenge, Zhang et al. [10] proposed a semi-autoencoder model. The semi-autoencoder does not require the dimensions of the input layer and the output layer to be consistent, and often the dimensionality of the input layer is greater than that of the output layer. The semi-autoencoder can also reconstruct part of the input data, such as the rating part. This model can flexibly utilize the additional attributes of the item and the user. However, due to the high cost of obtaining label knowledge and the increment awareness of user privacy protection, it is difficult to capture the additional information of user attributes.

To address these challenges, we propose a novel and simple model (Item-Agrec) that employs a semi-autoencoder to co-embed the attributes and the graph features of the items for rating prediction. The framework of our proposed Item-Agrec model is shown as Fig. 1. More specifially, we introduce a semi-autoencoder to learn the hidden nonlinear features of items for achieving a low computational cost, and the proposed Item-Agrec model can flexibly leverage additional information from different sources, such as attribute information, internal graph information, etc. Meanwhile, in order to solve the problem that the user’s attribute information is not easy to obtain, we introduce the Bipartite Network Embedding (BINE) [11] model to extract the internal graph features of the item from the user-item co-occurrence graph, and integrate them into the model for higher-level feature representations learning. Finally, we take the item’s graph features and attributes into consideration for improving the accuracy of recommendation. Our key contributions of this paper are summarized as follows:

  • A semi-autoencoder is employed to incorporate item attributes and graph features for exploring rich semantic representations or reconstructions, which can flexibly use attributes and content information from different sources with their vector representations for improving the performance of recommendation.

  • The graph features of the items are extracted and then integrated into the model for higher-level feature representation learning. For more sparse data sets, our model can effectively alleviate the problem of data sparsity and improve the accuracy of recommendations.

  • The experimental results conducted on several public real-world datasets demonstrate the effectiveness of the proposed method for capturing more powerful feature representations compared with the start-of-the-art methods.

The remainder of this paper is organized as follows. In Section 2, we briefly review the related work. We introduce the definition of the rating prediction problem in recommendation and related preliminary knowledge in Section 3, and then the model of Item-Agrec is proposed in Section 4. In Section 5, we give the experimental setup and the experimental results of three real-world data sets in detail. In Section 6, we summarize our proposed model and introduce the future work.

Section snippets

Related work

The objective of recommendation system is to estimate the user’s preferences for items and actively recommend items or services that the user may like [12], [13]. In the early research of recommendation system, collaborative filtering method is considered to be the most popular and widely used recommendation method [1]. Previous collaborative filtering methods can be divided into two categories, memory-based collaborative filtering and model-based one [14]. Memory-based methods are divided into

Preliminaries

Before we introduce our proposed model in detail, the preliminaries which applied in the method will be reviewed as follows. In this section, the commonly used notations in this paper are listed firstly, then the semi-autoencoder model for additional information combination and the BINE model for graph features learning are described in detail.

Item-Agrec

The whole framework of the proposed Item-Agrec is illustrated as Fig. 1, and the proposed model is divided into two main components. The first component is the graph features learning of the item with the BINE model, and the second one is a hybrid collaborative filtering model that predicts user-item ratings by learning item latent representations through the non-linear co-embeddings of their input features. In the following subsections we will discuss each component in details.

Experiments

In this section, we conduct extensive experiments on three real-world datasets to systemically demonstrate the effectiveness of our proposed framework. In the following, firstly, we introduce the details of recommendation datasets. Secondly, the compared benchmark methods are introduced in detail. Then the experiment results with observations of our proposed Item-Agrec and other competing methods are given. Finally, the main parameters of Item-Agrec are analyzed with certain dataset.

Conclusion

In this paper, we propose a novel and simple unsupervised deep learning framework for rating prediction, called Item-Agrec. It improves previous methods with a novel framework by incorporating item’s attributes information and graph features to optimize both feature representations and parameters of the learning model. In our framework, we use the BINE model to learn the graph feature vector of the item. The items rating vector, attributes feature vector, and graph feature vector are spliced

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.

Acknowledgements

This work is supported in part by the National Key Research and Development Program of China under grant 2016YFB1000901, the Natural Science Foundation of China under grants (61906060, 61906070, 62076085, 62076087, 91746209, 61876206), and the Program for Changjiang Scholas and Innovative Research Team in University (PCSIRT) of the Ministry of Education under grant IRT17R32.

BingBing Dong is a graduate student of Key Laboratory of Knowledge Engineering with Big Data (Hefei University of Technology) and School of Computer Science and Information Engineering at the Hefei University of Technology, Hefei, China. She received the BS degree from Anhui University of Finance and Economics. Her research interests are in data mining and recommendation system.

References (38)

  • J. Jiang et al.

    A fast deep autoencoder for high-dimensional and sparse matrices in recommender systems

    Neurocomputing

    (2020)
  • F. Zhuang et al.

    Representation learning via dual-autoencoder for recommendation

    Neural Networks

    (2017)
  • Z. Batmaz et al.

    A review on deep learning for recommender systems: challenges and remedies

    Artif. Intell. Rev.

    (2019)
  • Y. Koren

    Factorization meets the neighborhood: a multifaceted collaborative filtering model

  • X. Luo et al.

    Temporal pattern-aware qos prediction via biased non-negative latent factorization of tensors

    IEEE Trans. Cybern.

    (2019)
  • X. Luo et al.

    Position-transitional particle swarm optimization-incorporated latent factor analysis

    IEEE Trans. Knowl. Data Eng.

    (2020)
  • S. Sedhain et al.

    Autorec: Autoencoders meet collaborative filtering

  • Y. Wu et al.

    Collaborative denoising auto-encoders for top-n recommender systems

  • A. Rashed et al.

    Attribute-aware non-linear co-embeddings of graph features

  • S. Zhang, L. Yao, X. Xu, S. Wang, L. Zhu, Hybrid collaborative recommendation via semi-autoencoder, in: International...
  • M. Gao et al.

    Bine: Bipartite network embedding

  • G. Adomavicius et al.

    Toward the next generation of recommender systems: a survey of the state-of-the-art and possible extensions

    IEEE Trans. Knowl. Data Eng.

    (2005)
  • F. Ricci, L. Rokach, B. Shapira, Recommender systems: introduction and challenges, in: Recommender systems handbook....
  • F. Zhuang et al.

    Representation learning with pair-wise constraints for collaborative ranking

  • W. Yue et al.

    A hybrid model-and memory-based collaborative filtering algorithm for baseline data prediction of friedreich’s ataxia patients

    IEEE Trans. Industr. Inf.

    (2020)
  • X. Luo et al.

    An effective scheme for qos estimation via alternating direction method-based matrix factorization

    IEEE Trans. Serv. Comput.

    (2016)
  • X. He et al.

    Neural collaborative filtering

  • D. Kim et al.

    Convolutional matrix factorization for document context-aware recommendation

  • A. Van den Oord et al.

    Deep content-based music recommendation

    Adv. Neural Inf. Process. Syst.

    (2013)
  • Cited by (12)

    • Target-driven user preference transferring recommendation

      2024, Expert Systems with Applications
    • An improved hybrid semi-stacked autoencoder for item-features of recommendation system (iHSARS)

      2023, Indonesian Journal of Electrical Engineering and Computer Science
    View all citing articles on Scopus

    BingBing Dong is a graduate student of Key Laboratory of Knowledge Engineering with Big Data (Hefei University of Technology) and School of Computer Science and Information Engineering at the Hefei University of Technology, Hefei, China. She received the BS degree from Anhui University of Finance and Economics. Her research interests are in data mining and recommendation system.

    Yi Zhu is currently an assistant professor in the School of information Engineering, Yangzhou University, China. He received the BS degree from Anhui University, the MS degree from University of Science and Technology of China, and the PhD degree from Hefei University of Technology. His research interests are in data mining and knowledge engineering.

    Lei Li is an associate professor of Key Laboratory of Knowledge Engineering with Big Data (Hefei University of Technology), Ministry of Education, Hefei, China, Intelligent Interconnected Systems Laboratory of Anhui Province (Hefei University of Technology), Hefei, China, and School of Computer Science and Information Engineering, Hefei University of Technology, Hefei, China. He received his Bachelor’s degree in information and computational science from Jilin University, Changchun, China, in 2004, his Master’s degree in applied mathematics from the Memorial University of Newfoundland, St. John’s, Canada, in 2006, and his Ph.D. degree in computing from Macquarie University, Sydney, Australia, in 2012. He has published over 70 peer-reviewed papers in prestigious journals and top international conferences including IEEE Transactions on Cybernetics, IEEE Transactions on Services Computing, Knowledge-Based Systems, World Wide Web Journal, Pattern Recognition Letters, AAAI, IJCAI, ICDM, ICSOC and IEEE ICWS. His research interests include graph computing, social computing, data mining and intelligent computing. He is a senior member of IEEE.

    Xindong Wu is the President of Mininglamp Academy of Sciences, Mininglamp Technology, Beijing, China, and a Chang Jiang Scholar in the School of Computer Science and Information Engineering at the Hefei University of Technology, China. His research interests include data mining, knowledge-based systems, and web information exploration. He received his Ph.D. in artificial intelligence from the University of Edinburgh. He is the Steering Committee Chair of the IEEE International Conference on Data Mining (ICDM), the Editor-in-Chief of Knowledge and Information Systems, and Editor-in-Chief of the Springer book series, Advanced Information and Knowledge Processing (AIKP). He is Fellow of IEEE and the AAAS.

    View full text