Friendship prediction in social networks using developed extreme learning machine with Kernel reduction and probabilistic calculation

Abstract

The social network remains a highly dynamic object. Friendship prediction presents a significant problem in the research in network application in general and in social networking applications in particular. It involves analyzing an existing network graph and predicting more links inside the graph that were not identified before. Various models and approaches were developed for this purpose. Similarity-based models were used extensively, mainly they suffered from non-capability of handling the changing nature of the graph. Other models have supervised models that require training on labelled data. However, they need the extraction of many features to achieve satisfying performance. This work provides a novel implicit link prediction probabilistic reduced kernel extreme learning machine named ILP-PRKELM. Unlike the traditional supervised model of link prediction, ILP-PRKELM is attributed to the capability of achieving absolute accuracy with less number of features. Experimental results showed the superiority of ILP-PRKELM with an accomplished accuracy of 84.6 and 78.6 for Last.fm and Douban respectively, which is equivalent to 2% improved accuracy over the benchmarks.