An Experimental Evaluation of Link Prediction for Movie Suggestions Using Social Media Content

An Experimental Evaluation of Link Prediction for Movie Suggestions Using Social Media Content

Anu Taneja, Bhawna Gupta, Anuja Arora
Copyright: © 2018 |Pages: 28
DOI: 10.4018/978-1-5225-5097-6.ch011
(Individual Chapters)
No Current Special Offers


The enormous growth and dynamic nature of online social networks have emerged to new research directions that examine the social network analysis mechanisms. In this chapter, the authors have explored a novel technique of recommendation for social media and used well known social network analysis (SNA) mechanisms-link prediction. The initial impetus of this chapter is to provide general description, formal definition of the problem, its applications, state-of-art of various link prediction approaches in social media networks. Further, an experimental evaluation has been made to inspect the role of link prediction in real environment by employing basic common neighbor link prediction approach on IMDb data. To improve performance, weighted common neighbor link prediction (WCNLP) approach has been proposed. This exploits the prediction features to predict new links among users of IMDb. The evaluation shows how the inclusion of weight among the nodes offers high link prediction performance and opens further research directions.
Chapter Preview


Social Networks are the most popular way to model the interactions among users in the network-users can communicate among users through online social platforms like Twitter, Facebook etc to exchange information. The data in social networks can be considered in the form of links between objects. This can be modeled as graph where objects represent nodes and edges correspond to links. But social networks are not static; they are dynamic in nature as nodes and links that connect the nodes keeps on increasing which led to evolution of network i.e. nodes may enter or exit the network and interaction among the nodes represented by links may weaken or strengthen over time. This dynamic nature of social networks makes the task of link prediction a very challenging task in analysis of social networks. The links can be predicted using various social network analysis and mining techniques. Most network models capture global information (user profile) but do not capture local information (attended events, check-in status). Local information covers details about links to be formed and this type of information is more important for various applications such as event recommendation, next check-in location prediction etc. Link prediction approach helps in forming network of users on the basis of local information. In this chapter, author focuses on real environment problem which can be resolved through link prediction.

The dynamic and sparse nature of social networks makes difficult to predict links across a network. Though the task of link prediction is important as it forms the entire network, depict the relationship between individuals but is a highly challenging task. So, can more links be predicted on social networks? The research on this can be beneficial to assess the relationship among individuals and resonate the social behavior of individuals.

Although link prediction has been applied on wide variety of areas like bibliographic domain (Yu, 2012), molecular biology (Hammond, 2001; Setubal, 1997; Stamenkovic, 1989) criminal investigations (Chen, 2004; Huang, 2009), prediction of unknown information on protein networks (Link, 1997; Marcotte, 1999; Roy, 2010; Smith, 1996; Zuker, 1999) and recommender systems (Barbieri, 2014; Chen, 2005; Emanuelsson, 2000; Guns, 2014; Li, 2013; Liu, 2010; Tang, 2015) then also more effort is required in order to investigate link prediction area. While recommendation, the most popular and successful approach of recommendation is collaborative filtering (Ekstrand, 2011) which takes user-item rating matrix as input and generate recommendations by identifying similar users on the basis of rating patterns. But the main limitation with this approach is sparse nature of user-item rating matrix. To reduce the degree of sparsity, the recommendation problem is viewed as task of predicting links known as link prediction problem. Therefore the main motivation behind this work is to reduce sparsity as it leads to the formation of incomplete links among nodes than the number of links that should exist in the network using link prediction during recommendation and prediction. Another motivation is that generally people trust on friends more, so there is a need to find social patterns formed across user network and combine it with the network structure information.

Complete Chapter List

Search this Book: