ROBIN: A novel personal recommendation model based on information propagation

Highlights

• In this paper, we propose a novel personal recommendation model, called ROBIN.

• ROBIN employs Information Propagation and the EM method to implement the recommendation task.

• Experiment results show that ROBIN achieves good performance.

Abstract

With the rapid development of the Internet technology, we have now entered the era of information overloading. Recommendation System technology can recommend web resources or information to people based on his/her personal preference, and has gotten a great deal of attention and development in recent years. In this paper, by combining collaborative filtering technology and information propagation principle, we proposed ROBIN, a novel recommendation model. The ROBIN model achieves a good recommendation effect by propagating the relationship information between users and resources. Based on the ROBIN model, we designed and implemented tag recommendation algorithm named ROBIN-T. For evaluating our proposed method, we have conducted tag recommendation experiments on three real datasets and the results show that the ROBIN-T algorithm achieves good performance when compared with classical approaches.

Introduction

Because of the rapid development of the Internet technology, there are now so many information in the network that one person cannot look for needed information only depend on his/her own capacity. Because Recommendation System (Mahmood and Ricci, 2009, Resnick and Varian, 1997, Ricci et al., 2011) can recommend web resources or information to people automatically, it help people greatly reducing the cost of searching information. In recent years, recommendation system technology has gotten a great deal of attention and development, and has been used in more and more e-commerce websites.

Existing recommendation systems can be divided into three categories. The first one is non-personal recommendation system. This kind of recommendation system creates the same recommendation for every user with various kinds of ranking lists. Mostly it only uses some simple statistics methods. The second is semi-personal recommendation system. It recommends resources to user based on the resource he/she just look through. Techniques based on association rule mining are often used in such recommendation system. The last one is personal recommendation system. It automatic forms the behavior template for users by analyzing their history behaviors. personal recommendation system have been widely used in Web 2.0 applications such as Bibsonomy (http://bibsonomy.org) (sharing websites and publications), Flickr (http://www.flickr.com/) (sharing photos), Del.icio.us (http://delicious.com/) (sharing bookmarks), Movielens (http://movielens.umn.edu/) (sharing movies), etc.

The general procedure of personal recommendation consists of: (1) record user’s behaviors; generate personal behavior template by user’s history behaviors; (2) maintain behavior template; (3) recommend information to users based on their behavior template. Many technologies are proposed to solve the problem of personal recommendation (Resnick and Varian, 1997, Ricci et al., 2011).

In this paper, we propose a novel personal recommendation model named ROBIN, which is based on the information propagation. In the ROBIN model, a recommendation system is transformed to a bipartite graph, in which nodes of the graph represent users and resources respectively, and edge between a user and a resource represent the user’s potential action on the resource. A potential action is an action, such as tagging or browsing the resource, with a possibility which indicates that the opportunity that the action may be performed by a user on a resource.

The core problem of this paper is how to get a ROBIN model, especially the probability parameters, from the actual action network. For dealing with the problem, we propose an effectively EM algorithm to evaluate the probability parameters. In the algorithm, potential actions are propagated in the bipartite graph circularly. When the network becomes stable, the propagation processing stop, and then we get the ROBIN model.

When we know the potential action of a user from the ROBIN model, we can use it to predict the user’s behavior and get a good recommend result. To evaluate the effectiveness of the ROBIN model, we conduct experiments on three real datasets, which originate from Bibsonomy, Movielens, and Del.icio.us respectively. These experiments aim to personal tag recommendation task. Experimental results show that our method achieves better performance when compared with classical approaches.

The rest of the paper is organized as follows. Section 2 discusses related work. Section 3 presents the preliminaries of our work. In Section 4, we present the ROBIN model. In Section 5, we systematically develop an algorithm based on information propagation to obtain the probability parameters of the ROBIN model. Section 6 presents the experimental evaluation of the ROBIN model in tag recommendation task on three real datasets. Section 7 concludes this study and points out future works.