Ensemble Assisted Multi-Feature Learnt Social Media Link Prediction Model Using Machine Learning Techniques

Abstract

In this paper a robust consensus-based ensemble assisted multi-feature learnt social media link prediction model is developed. Unlike classical methods, a multi-level enhancement paradigm was considered where at first the focus was made on extracting maximum possible features depicting inter-node relationship for high accuracy of prediction. Considering robustness of the different feature sets, we extracted local, behavioural as well as topological features including Jaccard coefficient, cosine similarity, number of followers, intermediate followers, ADAR. The use of these all features as link-signifier strengthened the proposed link-prediction model to train over a large data and to ensure higher accuracy. Undeniably, the use of aforesaid multiple features-based approach could yield higher accuracy and reliability; however, at the cost of increased computation. To avoid it, different feature selection methods like rank sum test, cross-correlation, principal component analysis were applied. The use of these feature selection methods had dual intends; first to assess which type of features can have higher accuracy and second to reduce unwanted computation. This research revealed that cosine similarity-based features don’t have significant impact on eventual classification. On the contrary, cross-correlation and PCA based features had exhibited relatively higher accuracy (up to 97%). Once retrieving the set of suitable features, unlike standalone classifier based (two-class) prediction, we designed a novel consensus based ensemble learning model by using logistic regression, decision tree algorithm, deep-neuro computing algorithms (ANN-GD and ANN-LM with different hidden layers), which classified each node-pair as Linked or Not-Linked. Our proposed link-prediction model has exhibited link-prediction accuracy (98%), precision (0.93), recall (0.99), and F-Measure (0.97), which is higher than the other state-of-art machine learning methods.