| Home

Overview


Original Research

LINK PREDICTION ON MULTI ATTRIBUTE EXTRACTED SOCIAL NETWORK USING MACHINE LEARNING MODELS

UGRANADA CHANNABASAVA 1, and B K RAGHAVENDRA 2.

Vol 18, No 02 ( 2023 )   |  DOI: 10.17605/OSF.IO/96NDA   |   Author Affiliation: Department of Information Science & Engineering, Don Bosco Institute of Technology, Bengaluru, Karnataka, Affiliated to Visvesvaraya Technological University, Karnataka, India 1; Professor and Head, Department of Information Science & Engineering, Don Bosco Institute of Technology, Bengaluru, Karnataka, Affiliated to Visvesvaraya Technilogical University, Karnataka, India 2.   |   Licensing: CC 4.0   |   Pg no: 800-814   |   Published on: 15-02-2023

Abstract

An effective ensemble-based consensus-based multi-feature learned social media link prediction model is created in this research. Contrary to traditional methodologies, an improvement paradigm with many levels was taken into account. Where the initial emphasis was on extracting the most characteristics feasible that showed inter-node relationships for high prediction accuracy. We extracted local, Behavioural, as well as topological features, such as the Jaccard coefficient, cosine similarity, number of followers, intermediate followers, ADAR, shortest path, page rank, Katz coefficient index, hitting time of hops, and preferential attachment, taking into account the robustness of the various feature sets. The suggested link-prediction model was reinforced by using all of these attributes as link-signifiers, allowing it to be trained over larger datasets and with greater accuracy. Undoubtedly, using the aforementioned multiple features-based strategy might result in more accuracy and dependability, but at the expense of more computation. Different feature selection techniques, including the Gini index (GI), information gain (IG), PCA, and cross correlation (CC), were used to prevent it. These feature selection techniques were used with two goals in mind: first, to determine which types of features can have greater accuracy, and second, to minimise unnecessary computation. According to this study, cosine similarity-based characteristics don't significantly affect final categorization. In order to categorise each node-pair as Linked or Not-Linked, we developed a unique consensus-based ensemble learning model employing deep-neuro computing methods (ANN-LM with several hidden layers). Our suggested link-prediction model outperformed existing machine learning techniques in terms of link-prediction accuracy (98.7%), precision (0.95), recall (0.99), and F-Measure (0.93).


Keywords

Ensemble Learning Model, Feature Extraction, Social Network, Feature Selection