Personalized Recommendation Algorithm Research Based On Mass Diffusion

With the proliferation of mobile terminals and the rapid development of Internet technology,an explosion of data has been witnessed leading to the "information overload" problem,which is exacerbated by the accumulation with time and has also become a serious obstacle to the development of Internet technology.The information retrieval technique based on traditional search technology is no longer unable to meet the needs of users for information discovery.Personalized recommendation system predicts user's preference according to user's historical behavior information,provides personalized service for user,and actively recommends to users more information that is consistent with user's interest,which breaks the impasse that users are unable to get personalized information easily in the massive data.Personalized recommendation has become the main tool to assist people with information filtering and decision-making.So far,many researchers have proposed a variety of recommendation algorithms for different application scenarios.Among them,by virtue of good recommendation performance and adaptability,the recommendation algorithms based on mass diffusion has attracted the widespread attention of scholars in recent years.In this thesis,we aim to deeply analyze the fundamental and problems existing in current recommendation algorithms based on mass diffusion,and put forward two improved ones for better and more comprehensive performance.The main innovation of this thesis are as follows:(1)This thesis proposes a personalized recommendation algorithm based on preferential bidirectional mass diffusion.Most of the algorithms based on mass diffusion focus on unidirectional mass diffusion from collected objects to uncollected objects,leading to the risk of similarity estimation deviation due to the asymmetry of similarity.In addition,they are biased towards recommending popular objects which causes popularity bias problem.This thesis proposes a personalized recommendation algorithm based on preferential bidirectional mass diffusion by considering both the positive similarity from collected objects to uncollected objects and the reverse similarity from the uncollected objects to collected objects.What's more,the weight of popular objects in bidirectional similarity is penalized to alleviates popularity bias.Experiments are evaluated on three real datasets(Movielens,Amazon and RYM)by 10-fold cross validation.Results show that the proposed recommendation algorithm based on preferential bidirectional mass diffusion can effectively correct the similarity estimation deviation,alleviate the popularity bias problem and greatly enhance the recommendation performance.(2)This thesis proposes a personalized recommendation algorithm based on suppressing excessive mass diffusion.Realistic recommended network structure is complex,and there exists excessive diffusion phenomenon in recommendation algorithms based on mass diffusion in bipartite network which simulate the process of resource redistribution process.It leads to an overestimate of user's preferences in some items,which serious affects the performance of recommendation.Against two main problems(popularity bias and redundant similarity)caused by excessive diffusion,this thesis penalizes the popular objects and user's heterogeneity of choice to suppress popularity bias problem,and then leverage the second-order similarity to reduce similarity redundancy.Experiments are evaluated on three real benchmark datasets(MovieLens,Amazon,and RYM)by 10-fold cross validation shows that the proposed algorithm can more effectively suppress the problem of excessive mass diffusion,and improve recommendation accuracy,diversity,and novelty.