Research And Implementation Of Probability PPV Model For Influence Maximization Problem

With the growth of smart devices,social network services(like Tao Bao item recommendation,Dianping location-based recommedation,and Wangyi Music recommendation,etc.) become popular.People used to trust information coming from their close relatives and friends.Because of the popularity of social network,online "viral marketing" which focuses on influence maximization problem has awakened a great interest.As viral marketing takes users' social gragh into consideration,it helps to tackle the problem of matrix sparsity.Besides,presicion of recommendation results is more accurate.Thus,influence maximization problem on social network is worth further studies.Our research proposed a PPV probability model based on IM problem which effectively learn influence probabilities in social network,and personalize influence for each node pair.By clustering user groups and analyzing similarity of users from both offline action log and online social network topological structure,we estimate reliable parameters of our probability model,tell user's different influence on its neighbors because of different features,and improve probability learning procedure of influence maximization problem.Experiments show that our approach outperforms the state-of-the-art algorithm.Details are as follows:1.Reliable influence graph extraction:Existing researches focus on topological structure anaylises and neglect the importace of nodes themselves.Our research provides an analysable and reliable influence graph G with theoretical guarantee.2.Influence probability learning:Traditional probability learning is not theoretical guaranteed and use unitary methods.Our research analyse topological structure to explain states transition and calculate transition probability,thus to give reasons of the spread of typical event.3.Prototype system and testing:In this paper,we use Gowalla data set to conduct experiments,and compare MAE and RMSE to evaluate results.Experiments show that our approach outperforms the state-of-the-art algorithm.