Search

For niche recommendation, information filtering has attracted much attention from various fileds. Especially, mass diffusion based models behave prominently. Nevertheless, these models intrinsically suffer superfluous diffusion, transferring redundant preferences to the object and damaging the accuracy, diversity, and personalization of recommendation. Besides, we discover that the symmetrical diffusion can effectively improve recommendation performances. Thus, we assume that the superfluous diffusion should be symmetrically punished. Hence, we propose a symmetrical punishment model on superfluous diffusion for accurate information recommendation. Extensive experiments on two data sets N e t f l i x and M o v i e l e n s show that our proposed model outperforms mainstream indices remarkably in accuracy, diversity and personalization. [ABSTRACT FROM AUTHOR]

In physical dynamics, mass diffusion theory has been applied to design effective information filtering models on bipartite network. In previous works, researchers unilaterally believe objects’ similarities are determined by single directional mass diffusion from the collected object to the uncollected, meanwhile, inadvertently ignore adverse influence of diffusion overload. It in some extent veils the essence of diffusion in physical dynamics and hurts the recommendation accuracy and diversity. After delicate investigation, we argue that symmetrical diffusion effectively discloses essence of mass diffusion, and high diffusion overload should be published. Accordingly, in this paper, we propose an symmetrical and overload penalized diffusion based model (SOPD), which shows excellent performances in extensive experiments on benchmark datasets Movielens and Netflix . [ABSTRACT FROM AUTHOR]

Recently, in complex network, link prediction has brought a surge of researches, among which similarity based link prediction outstandingly gains considerable success, especially similarity in terms of paths. In investigation of paths based similarity, we find that the effective influence of endpoints and strong connectivity make paths contribute more similarity between two unconnected endpoints, leading to a more accurate link prediction. Accordingly, we propose a so-called effective path index (EP) in this paper to leverage effective influence of endpoints and strong connectivity in similarity calculation. For demonstrating excellence of our index, the comparisons with six mainstream indices are performed on experiments in 15 real datasets and results show a great improvement of performance via our index. [ABSTRACT FROM AUTHOR]

Abstract Identifying the influential spreaders of information spreading dynamics is a hot topic in the field of network science. To identify the influential spreaders, most previous studies were based on the global information of the network. In this paper, we propose a strategy for identifying the influential spreaders from a randomly selected initial-seed node. The seeds are connected as a chain, and are localized to the initial-seed. In our proposed preferentially random walk based influential spreaders identifying strategy, the walker's movement is adjusted by neighbors' degrees. The seeds are those nodes that the walker ever visited. Through extensive numerical simulations on artificial networks and four real-world networks, we find that selecting large degree nodes preferentially is more likely to find the most influential spreaders. The outbreak threshold decreases when preferentially select hubs. Our results shed some light into identifying the most localized influential spreaders. Highlights • Proposing a strategy for identifying the localized influential spreaders from a randomly selected initial-seed node. • Selecting large degree nodes preferentially is more likely to find the most influential spreaders. • The effectiveness of the strategy is verified in both artificial and real-world networks. [ABSTRACT FROM AUTHOR]

Recommendation system provides a promising way to alleviate the dilemma of information overload. In physical dynamics, mass diffusion has been used to design effective recommendation algorithms on bipartite network. However, most of the previous studies focus overwhelmingly on unidirectional mass diffusion from collected objects to uncollected objects, while overlooking the opposite direction, leading to the risk of similarity estimation deviation and performance degradation. In addition, they are biased towards recommending popular objects which will not necessarily promote the accuracy but make the recommendation lack diversity and novelty that indeed contribute to the vitality of the system. To overcome the aforementioned disadvantages, we propose a preferential bidirectional mass diffusion (PBMD) algorithm by penalizing the weight of popular objects in bidirectional diffusion. Experiments are evaluated on three benchmark datasets (Movielens, Netflix and Amazon) by 10-fold cross validation, and results indicate that PBMD remarkably outperforms the mainstream methods in accuracy, diversity and novelty. [ABSTRACT FROM AUTHOR]

People normally show different relation intensity and adoption behaviors, which can affect information spreading on social networks. To describe the two phenomena, we consider edge weight distribution and adoption thresholds heterogeneity on the networks separately, where the adoption thresholds of nodes obey truncated gaussian distribution. Therefore, we propose a double-layer network model with edge weight distribution to explore the impact of heterogeneous adoption thresholds of nodes on information spreading. And we also develop an edge-based compartmental theory to analyze the information spreading mechanism. Through numerical simulations and theoretical analysis, we find that increasing the values of mean can suppress information spreading, and the impact of the values of standard deviation on information spreading depends on the values of mean. Specifically, for a small value of mean, the size of information spreading will decrease with the increase of the value of standard deviation. Meanwhile, when the values of standard deviation are fixed, the size of information spreading will increase continuously with the increase of transmission probability. For a large value of mean, the size of information spreading will increase first and then decrease with the increase of the value of standard deviation. In addition, this paper shows that the theoretical method agrees with the numerical simulations in the experimental results. • The double-layer propagation model with edge weight distribution is structured. • The impact of adoption threshold heterogeneity is studied on the model. • The adoption thresholds of nodes obeys a truncated Gaussian distribution. • The phase transition depends on the mean and the standard deviation. [ABSTRACT FROM AUTHOR]

In complex networks, topological similarity based link prediction promotes the development of network science. Traditional researches consider that two unconnected endpoints have possibility to make a link if they possess large influence respectively. However, through profound investigations, we find that one endpoint with large influence also can attract other endpoints around. The phenomenon reveals that the mutual attractions between two unconnected endpoints depend on their combined influence instead of only single influence. Furthermore, previous researches pay more attention to the influence of endpoints with only degree considered. However, in quasi-local paths, we discover that synthesizing degree and H-index can more reliably capture the endpoints with great and extensive maximum connected subgraph, which can more possibly attract other unconnected endpoints. To sum up, we propose a model named combined hybrid influence connectivity index (CHIC) in this paper to explore its role on similarity based link prediction. The comparisons with eight mainstream indices are performed on experiments in twelve real data sets and the results show an improvement of prediction performance via CHIC index. • The links between two unconnected endpoints depend on their combined influence. • Synthesizing degree and H-index can represent the influence of endpoints accurately. • The similarity of two endpoints depends on the connectivity of quasi-local paths. [ABSTRACT FROM AUTHOR]