Affiliations:
School of Information Science and Engineering, Central South University, Changsha, Hunan, China
Correspondence:
[*]
Corresponding author. Xiaoheng Deng, School of Information Science and Engineering, Central South University, Changsha, Hunan, China. Tel.: +86 13508482734; Fax: +86 073185536504; E-mail: dxh@csu.edu.cn.
Note: [1] Preliminary results of this work have been presented in IEEE FSKD’15 [26]. The new contributions of this manuscript include an improved design of the user affinity in (8) associated with the conditional probability and the disposition of extremely cases, an enhanced combination of user static influence procedure (as shown in 10) by considering the the range properties of measurement in section 3, an in-depth analysis and more comprehensive evaluations by simulating different network scenario with another dataset (as shown in Table. 2) to demonstrate the working performance of CD-NF model and GNF algorithm (Figs. 2, 3) in Section 5, and enriched related work by a detailed comparison of the proposed model and algorithm with the existing ones in Section 2.
Abstract: Influence maximization is a problem of identifying a small set of highly influential individuals such that obtaining the maximum value of influence spread in social networks. How to evaluate the influence is essential to solve the influence maximization problem. Meanwhile, finding out influence propagation paths is one of key factors in the assessment of influence spread. However, since nodes’ degrees are utilized by most of existent models and algorithms to estimate the activation probabilities on edges, node features are always ignored in the evaluation of influence ability for different users. In this paper, besides the node features, the Credit Distribution (CD) model is extended to incorporate the time-critical aspect of influence in online social networks. After assigning credit along with the action propagation paths, we pick up the node which has maximal marginal gain in each iteration to form the seed set. The experiments we performed on real datasets demonstrate that our approach is efficient and reasonable for identifying seed nodes, and the influence spread prediction by our approach is more accurate than that of original method which disregards node features in the influence evaluation and diffusion process.
