Elsevier

Information Sciences

Volume 447, June 2018, Pages 140-156
Information Sciences

Multi-Sided recommendation based on social tensor factorization

https://doi.org/10.1016/j.ins.2018.03.019Get rights and content

Abstract

Tensor factorization has been applied in recommender systems to discover latent factors between multidimensional data such as time, place, and social context. However, tensor-based recommender systems still encounter with several problems such as sparsity, cold-start, and so on. In this paper, we introduce the new model social tensor to propose a tensor-based recommendation with a social relationship to deal with the existing problems. In addition, an adaptive method is presented to adjust the range of the social network for an active user. To evaluate our method, we conducted several experiments in the movie domain. The results indicate the ability of our method to improve the recommendation performance, even in the case of a new user. Particularly, the proposed method conducts the regeneration and factorization of the tensor in real time. Furthermore, our approach recommends not only a single item, but also the multi-factors for the item such as social, temporal, and spatial contexts.

Introduction

There is a large amount of various information on the web which has provided more freedom of choices to users. While it causes the users to encounter difficult choice, which is interest of them [12], [29]. In this regard, context-based recommender systems have been studied to give useful item for users by using various information [16]. The systems that are typically used are the following four contexts among the “Five Ws” concept: user, item, and temporal and spatial contexts. Most of the systems, however, only partially consider the contextual factors (e.g., “whowhatwhere” and “whowhatwhen”) [28], [41]. Thus, a number of studies have tried to apply the four factors into their recommender systems [31], [44].

Over the last two decades, tensor factorization has been proposed to discover latent factors in multidimensional data that comprise at least three dimensions. Also, several researchers who have tried to apply this technique in the context-based-recommendation process have shown that it is powerful for recommender systems with the multidimensional data [32], [39], [45], [47]. However, regarding these approaches, a number of problems (e.g., sparsity, cold-start, and re-generation) have arisen [15]. For instance, it is necessary to previously calculate an approximate tensor before recommendation, since a large amount of time is required to calculate the tensor. With respect to the situation, it is difficult to immediately reflect a new user (item) into tensor model [46]. Although methods have been proposed to solve the problem by applying the new user (item) in tensor model which is already approximated instead of the tensor generation, new users (items) are still reflected for only part of tensor model [9]. Therefore, it is important to generate a tensor model and to approximate it with short time.

To deal with the problems, the social relationships of an active user are used in this study for the tensor generation and factorization. To our best knowledge, the existing studies have taken the social factor into account for only the tensor factorization as a regularization term [2], [40], [45]. In this regard, the introduction of a new model, the “social tensor,” is presented here, along with a proposed recommendation method that can consider the five aspects of user, item, temporal context, spatial context, and social context.

The proposed method is divided into four steps. The first step is the construction of a tensor model that is based on a social network of active users. In the second step, the tensor factorization is conducted to obtain a suitable context for the users. The third step recommends a list of items based on the similarity between the candidate set and the movies in histories of the users. Lastly, the proposed method learns the range of the social network based on the user feedback.

Recently, the double-sided recommendation (DSR) has been introduced to provide both items and users for the improvement of the recommendation quality [19], [36]. The method name Multi-sided Recommendation based on Social Tensor (MSRST) was derived since this approach recommends multiple factors as follows: a movie as the item, two users (i.e., an active and another user for the watching of the recommended item together), time as the temporal context, and place as the spatial context.

The proposed approach brings a number of contributions to the area of the tensor-based recommendation, as follows:

  • introduction of the new social-tensor model for its (re)generation and factorization in real time, and

  • provision of multifactors (i.e., temporal context, spatial context, social context) for the item recommendation.

The rest of this paper is organized as follows. Section 2 reviews the previous work that relate to the proposed approach. In Section 3, an introduction of the proposed scenario and research questions are presented. The notations and definitions of the proposed social-tensor model are described in Section 4. In Section 5, the proposal of a learning method for the social tensor and the proposed recommendation algorithm are presented. The experiment results are presented and discussed in Section 6. Lastly, in Section 7, the conclusion is provided and a future work is introduced.

Section snippets

Context-based recommendation

Recently, the opportunity to acquire contextual data from users arose due to the advance of mobile technologies. In this regard, many researchers have tried to use the data for the design of a recommender system [4], [7], [8], [16], [21], [49]. According to the Five Ws concept, the contextual data has been separated into the four types of who, when, where, and what. To consider the context in the recommendation, several studies have imposed the four types as various formats like who-where-what,

Scenario and research questions

To describe the motivation of this study and to raise the research questions, a simple movie-domain scenario is established for a case study. Fig. 1 shows the histories of John with Mary and James that gives six contexts to John in terms of the temporal and spatial contexts.

Based on the histories, it is possible to discover the interdependent phenomena, as follows:

  • -

    when John is at home at midnight, he mainly enjoys watching movies alone.

  • -

    When John wants to watch a movie with Mary, he usually

Social tensor

In this section, a description of the definitions for the new social-tensor model is provided, along with an explanation of the construction, factorization, and data scheme of the proposed tensor.

Social tensor-based recommendation

In this section, the description of a MSRST algorithm is provided. Our method is divided into the following four steps: (i) construction of a tensor model based on a social network of active users, (ii) factorization of the tensor to obtain a suitable item context, (iii) recommendation of a list of items using the context, and (iv) learning the social-network multi-hop range based on the user feedback.

Experiments and discussion

In this section, the experiments are implemented in terms of the following three aspects: (i) number of multi-hops, (ii) performance of the proposed method, and (iii) the multi-sided recommendation. Firstly, an evaluation of the proposed method was conducted with respect to the following three factors: default value DH, parameter λ, and parameter μ. In the second aspect, a comparison of our method with several recommender systems in terms of the sparsity, the cold-start problem, and the

Conclusion and future work

Although many researches that enable a tensor in the recommender system have been performed, the corresponding methods suffer from the sparsity and cold-start problems. To alleviate these problems, a novel “social tensor” model is proposed here, and it is applied for a tensor-based recommendation through an adaptive method to adjust the range of the social network for an active user.

A number of experiments were conducted using a real-world dataset, which was gathered from the MyMovieHistory for

Acknowledgment

This work was supported by the National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIP) (NRF-2017R1A2B4010774).

References (49)

  • C. Zheng et al.

    CMPTF: contextual modeling probabilistic tensor factorization for recommender systems

    Neurocomputing

    (2016)
  • X. Zheng et al.

    Topic tensor factorization for recommender system

    Inf. Sci. (Ny)

    (2016)
  • H. Bagci et al.

    Context-aware friend recommendation for location based social networks using random walk

  • P. Bhargava et al.

    Who, what, when, and where: multi-dimensional collaborative recommendations using tensor factorization on sparse user-generated data

  • C. Biancalana et al.

    An approach to social recommendation for context-aware mobile services

    ACM Trans. Intell. Syst. Technol. (TIST)

    (2013)
  • J.S. Breese et al.

    Empirical analysis of predictive algorithms for collaborative filtering

  • Y. Cai et al.

    Low-order tensor decompositions for social tagging recommendation

  • R. Colomo-Palacios et al.

    Towards a social and context-aware mobile recommendation system for tourism

    Pervasive Mob. Comput.

    (2016)
  • E. Frolov et al.

    Tensor methods and recommender systems

    Wiley Interdiscip. Rev. Data Mining Knowl. Disc.

    (2017)
  • M. Hong et al.

    Mymoviehistory: social recommender system by discovering social affinities among users

    Cybern. Syst.

    (2016)
  • M. Hong et al.

    GRSAT: a novel method on group recommendation by social affinity and trustworthiness

    Cybern. Syst.

    (2017)
  • B. Hu et al.

    Spatial topic modeling in online social media for location recommendation

  • N. Ifada et al.

    Tensor-based item recommendation using probabilistic ranking in social tagging systems

  • A. Karatzoglou et al.

    Multiverse recommendation: n-dimensional tensor factorization for context-aware collaborative filtering

  • Cited by (0)

    View full text