Elsevier

Computer Communications

Volume 157, 1 May 2020, Pages 221-231
Computer Communications

User recommendation method based on joint probability matrix decomposition in CPS networks

https://doi.org/10.1016/j.comcom.2020.03.044Get rights and content

Abstract

In recent years, with the rapid development of the Internet, various virtual communities continue to emerge, and the phenomenon of user groups working together is gradually increasing. People begin to pay more attention to group-oriented recommendation. Most of existing group recommendation methods are improved on the memory-based collaborative filtering recommendation method, or think that the members of the group are independent of each other, ignoring the impact of association among the members of the group on the results of group recommendation. In this paper, a group recommendation method based on joint probability matrix decomposition is proposed to better model the group recommendation problem. Firstly, the user-plus-person group information is used to calculate the correlation between users. Secondly, the user correlation matrix is integrated into the process of probability matrix decomposition to get the individual prediction score. Finally, the group-to-item prediction score is obtained by using the common synthesis strategy in group-oriented recommendation problem. Furthermore, the proposed method is compared with existing group recommendation methods. Experiments on CiteULike dataset show that the proposed method achieves better recommendation results in accuracy, recall and other evaluation indicators.

Introduction

With the development of the information age, the major social networks are becoming more and more popular. While users use the network, many users generate content independently every day. It is difficult for users to find the information they are interested in from the huge amount of information. Although search engine is a way to solve this problem, it can only search for a specific information and cannot provide personalized active information service for every user [1]. As a result, recommendation system came into being to solve the problem of information overload in the context of the Internet era. Recommendation system is mainly based on the relevant data generated by users and searches interested information for users in a positive way. It transforms the original targeted information search problem into a more user-oriented information discovery problem. In recent years, it has been widely used in many fields such as information retrieval, e-commerce and online advertising, and is becoming a hotspot of current research [2], [3], [4].

Traditional recommendation systems focus more on generating recommendation lists for individual users. However, with the development of the Internet and social networks, the major social networks gradually appear group functions, such as Douban community has a variety of music groups, film groups and reading groups [5]. With the emergence of various groups on social networks, more and more users choose to join groups according to their interests. Users like to participate in certain activities with friends in the same group. These groups can enhance the communication among users, improve the information sharing among users, and reduce the time for users to search for items of interest [6]. It is noteworthy that the important role played by groups in the recommendation process is often neglected. Because users usually find interesting items according to experts’ comments or their own opinions, if the recommendation system directly generates suggestions to the user’s group, it will simplify the user’s search process to a certain extent. Therefore, group-oriented recommendation has become a new challenge.

At present, some researchers have done some research on group recommendation. By synthesizing individual users’ preferences or synthesizing individual users’ recommendation results, the final group recommendation results can be obtained [7]. Collaborative filtering is widely used in recommendation for individual users because it has been proved to be an effective recommendation method. Therefore, researchers also use collaborative filtering to solve group recommendation problems. For example, Kim and other scholars proposed a two-stage hybrid recommendation method. First, they used keyword technology to find the nearest neighbors, then used collaborative filtering method to generate recommendation lists, and finally deleted unrelated books from the draft recommendation list [8]. Some researchers used collaborative filtering to generate group recommendation results, and compared the differences of group recommendation results synthesized in four stages of collaborative filtering [7]. Other scholars proposed a hybrid method which combines collaborative filtering and content-based methods to recommend groups. Firstly, collaborative filtering recommendation list and content-based recommendation list are generated for each user. Then, group conflicts are solved by using fusion strategy. Finally, the collaborative filtering recommendation list is reordered by content-based recommendation list [9].

However, the above group recommendation methods still have some shortcomings. On the one hand, traditional memory-based collaborative filtering methods, such as UserKNN and ItemKNN, are facing the problem of data sparseness and weak scalability. On the other hand, most existing research methods do not make good use of the interaction relationship among group members in group recommendation, which makes the results of group recommendation deviate from the preferences of actual group members. To solve the problem of data sparsity, a model-based collaborative filtering method, matrix decomposition, can effectively alleviate this problem by using the idea of dimensionality reduction [10]. However, only a few researchers have introduced matrix decomposition into group recommendation. Secondly, existing group recommendation methods assume that members are independent of each other, while the group-based membership as an important information is rarely considered.

Based on the above two problems, this paper proposes a group recommendation method based on joint probability matrix decomposition. Firstly, the matrix decomposition method is introduced as the basic model of the recommendation framework. Secondly, the user’s interest preference information of different groups is integrated into the process of matrix decomposition, and the individual prediction score is generated by the joint probability matrix decomposition method. Finally, the synthesis strategy of group recommendation is selected to synthesize the individual prediction score of each group, and the result of group recommendation is generated. In this paper, CiteUlike dataset is used. The experimental results show that the proposed method achieves good results and improves the accuracy of group recommendation.

The rest of this paper is organized as follows: in Section 2, we introduce related works and point out the limitations of some previous works. In Section 3, We give the definition of related problems. In Section 4, we give the Group Recommendation Method Based on Joint Probability Matrix Decomposition. In Section 5, we conduct experiments to evaluate our approach. In Section 6, we Analysis and discussion of results. Finally, Section 7 concludes this paper.

Section snippets

Collaborative filtering method

Collaborative filtering is the most commonly used recommendation technology in recommendation system. It can be divided into two main categories: memory-based method and model-based method [11].

Memory-based collaborative filtering method finds neighbors with high similarity between users or projects through users’ historical information. According to the comprehensive evaluation of neighbors’ projects, users’ preferences for recommended projects are predicted [10], [11], [12]. In contrast, the

Question definition

This paper predicts the preferences of groups according to the preferences of users and the information of users joining groups. The formal definition of this problem is given as follows. Firstly, N users in the system are represented by U=u1,u2,,ui,,uN, M items are represented by V=υ1,υ2,,υj,,υM, R=Ri,jN×M refers to the matrix composed of the preference values of ui to υj. In addition, with respect to group information, L groups in the system are represented by G=g1,g2,,gl,,gL, matrixes

Group recommendation framework based on joint probabilistic matrix decomposition

In this study, matrix decomposition method is introduced in the process of group recommendation, and based on it, the relevant characteristics of group are fully considered. The method framework presented in this paper is shown in Fig. 3.

As can be seen from Fig. 3, the method proposed in this paper mainly includes the following steps:

(1) Measuring user relevance. The interrelatedness among users is acquired by the information contained by groups and users. Among them, the Group owned by users,

Data set

The data of this study comes from CiteUlike website, a research social networking website that can help researchers enhance academic exchanges. Users on this website can collect interesting papers while browsing and reading articles. At the same time, the website allows users to create research groups or interest groups, invite other users in the same field or hobby to join and share relevant information. Research and exchange. Therefore, the information contained in CiteUlike website coincides

Experimental results

The experimental results of this study are shown in Table 2, in which the number of input, input and recommendation is better than the number in the table.

In the results shown in Table 2, UPMF_AVG, UPMF_LM andUPMF_MP are the group recommendation methods proposed in this paper. From Table 2, we can see that: firstly, the three methods mentioned in this paper all perform better under several evaluation indexes, which further shows the effectiveness of the proposed group recommendation method in

Summary and future works

In this study, a group recommendation method based on joint probability matrix decomposition (JPMD) considering group information is proposed. Firstly, considering the group information to measure user correlation and combining the joint probability matrix decomposition method, the individual score is predicted. Secondly, in the stage of group rating, three fusion methods commonly used in group recommendation are selected to synthesize, and finally a group recommendation list is generated for

CRediT authorship contribution statement

Zhe Yao: Data curation, Formal analysis, Project administration, Software, Validation, Visualization, Writing - original draft, Writing - review & editing. Kun Gao: Conceptualization, Funding acquisition, Investigation, Methodology, Resources, 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.

References (29)

  • BerkovskyS. et al.

    Group-based recipe recommendations: analysis of data aggregation strategies

  • CarulloG. et al.

    A triadic closure and homophily-based recommendation system for online social networks

    World Wide Web

    (2015)
  • SalehiM. et al.

    Personalized recommendation of learning material using sequential pattern mining and attribute based collaborative filtering

    Educ. Inform. Technol.

    (2014)
  • KaurS. et al.

    Recommendation generation using typicality based collaborative filtering

  • Cited by (0)

    View full text