Loading...

Link Prediction in Social Networks Using the Diffusion Network Characteristics

Hossein Nazer, Tahora | 2013

1240 Viewed
  1. Type of Document: M.Sc. Thesis
  2. Language: Farsi
  3. Document No: 44817 (19)
  4. University: Sharif University of Technology
  5. Department: Computer Engineering
  6. Advisor(s): Rabiee, Hamid Reza
  7. Abstract:
  8. Given a snapshot of a network, link preditction methods try to infer future intractions be-tween its nodes. These methods may be used in either analyzing current state of the network or predicting future links of it. Link prediction techniques have many applications among which we can mention recommendation systems. These systems are implemented for com-mercial reasons or preventing user confusion in huge amount of information available.A new perspective toward link prediction is based on supervised random walk. In such methods, a random walker starts from a node in the network and randomly traverses to one of the current node’s neighbours with a probability proportional to the chosen link’s weight.In another category, content based methods use nodes’s properties and profiles as inputs of a binary classifier. Based on the likelihood of construction, these methods classify future links intoprobableandnot-probableones. Although algorithms based on supervised randomwalk out perform other proposed methods, they are based on the assumption that social network topology is available which seems to be impossible in some situations.In order to overcome the mentioned challenge, we aim to introduce a fast and accurate algorithm for link prediction in social networks using characteristics of the overlaying dif-fusion network; the network in which news, ideas, or disease spread through the underlying network. Due to the the relation between the social and diffusion network, the later may be used to recover unobserved links in the former. The social network investigated in the pro-posed method which includes users and items is directed and weighted. The links weights are determined based on timing properties of diffusion network and ages of items.We consider not having any information from the underlying social links, and try to re- cover and weight probable ones. To do so we use infection times of user sincasca des available to the proposed method. Moreoverwe weight collaboration links basedon items’age, which has been ignored in most of the link prediction methods. Finally we use a method based on supervised random walk to rank items in respect to a target user. Highly scored items may be recommended to the target user as future links. We will show that our algorithm “DiffRank”outperforms state-of-the-art prediction methods in the same field when we have incomplete information of the underlying network
  9. Keywords:
  10. Diffusion Network ; Social Networks ; Link Prediction ; Supervised Random Walk ; Recommender System

 Digital Object List