Diffusion-based Recommendation Algorithms On Coupled Social Networks

The rapid development of the Internet has brought a great impact to our daily life and greatly changed the way we get information.It provides us with abundant online content,so we need to spend a lot of time to browse and find the required information in the massive information,which is often referred to as "information overload" problem.In order to overcome the "information overload" problem,a lot of information filtering approaches was proposed by researchers,the most famous of which is the search engine and recommender system.The search engine can return the most relevant web pages based on the keywords provided by the user.Although search engine has been widely used,but it still has two obvious shortcomings:First of all,it is difficult for an unexperienced user to provide accurate search keywords;second,the search engine always returns identical search results for a given keyword.Recommender systems leverage user's history selection information and evaluation information to help the target user find the interested and useful information associated with it.Recommender systems do not require user to enter the keyword and returns the personalized recommendation result,which overcomes the shortcomings of the search engine.At present,recommendation algorithm has been widely used in E-commerce,computing advertising,online education and online medical.It has brought us great commercial value.In most practical applications,user's historical behavior information is very sparse.Due to the lack of sufficient historical data,recommender system is often difficult to give accurate recommendations for the target user,which is often referred to as "data sparseness problem",it has been troubled recommender systems for years.Social influence plays a very important role in product marketing,but the traditional recommender systems rarely incorporate it into the recommendation framework.With the rapid development of online social networks,social-based recommendations have become an important and effective way to recommend new items to users.The existing recommendation algorithms based on social network have their own advantages and disadvantages,but their common characteristic is that one needs additional information such as ratings or social tags,but the user's ratings or tags information are difficult to obtain,and the quality of these information cannot be guaranteed.In recent years,the recommendation algorithms based on bipartite networks draw more and more attention from researchers because of they only focus on the user's choice of items but not tags or ratings.These algorithms,based on the idea of mass diffusion and heat conduction,abstracts the data from user's historical information and incorporate it into complex network model,which is lower than the classical cooperative filtering algorithm in the computing complexity,and has good scalability.Aim at the problems and challenges of the existing approaches,this paper makes an in-depth study on the recommendation algorithms by using the diffusion process to integrate social network information into recommender systems.This paper investigates mass diffusion on coupled social networks,trust-propagation-aware mass diffusion on coupled social networks,and the recommendation algorithm which incorporates social network information into recommendation framework and mixes mass difEusion and heat conduction in the recommendation process.The main contributions of this paper are as follows:(1)A mass diffusion recommendation algorithm which integrates social network information into recommender systems was proposed.The algorithm uses the mass diffusion process to integrate the social network information into the recommendation process,control the participation of social network information through a regulation parameter.Experimental results show that the social network information plays an extremely important role in the recommendation process when the user chooses few items.In the case of "data sparseness problem",the recommendation accuracy outperforms the traditional mass diffusion and hybrid algorithm of mass diffusion and heat conduction.It shows that the algorithm is especially applicable to the two-layer coupled network,whose user-item bipartite network is relatively sparse and social network is relatively dense.The experimental results also show that the algorithm is a very good alternative to the global ranking method in the cold start period.(2)The algorithm utilizes the mass diffusion process to integrate the social network information organically into the recommendation process,not only using the selection information of the target user's direct friend,but also the selection information of the second-order friend of the target user.The experimental results show that the recommendation accuracy of the algorithm is much higher than that of the traditional mass diffusion algorithm when the target user chooses fewer items.And the recommendation diversity of traditional mass diffusion algorithm can be improved when the target user chooses lots of items.(3)A recommendation algorithm,which not only incorporates social information into the recommendation framework but also mixes mass diffusion and heat conduction in the recommendation process,was developed.Traditional mass diffusion method,heat conduction method,hybrid method and the mass diffusion algorithm on coupled social networks which proposed in this paper are the special cases of this algorithm.The experimental results show that the proposed algorithm is superior to mass diffusion on coupled social networks which proposed in this paper due to the mixture of mass diffusion and heat conduction in recommendation process.Recommendation accuracy of this algorithm has improved compared with the hybrid method.