We gratefully acknowledge support from
the Simons Foundation and member institutions.
Full-text links:

Download:

Current browse context:

cs.SI

Change to browse by:

cs

References & Citations

DBLP - CS Bibliography

Bookmark

(what is this?)
CiteULike logo BibSonomy logo Mendeley logo del.icio.us logo Digg logo Reddit logo ScienceWISE logo

Computer Science > Social and Information Networks

Title: Link Prediction Using Supervised Machine Learning based on Aggregated and Topological Features

Abstract: Link prediction is an important task in social network analysis. There are different characteristics (features) in a social network that can be used for link prediction. In this paper, we evaluate the effectiveness of aggregated features and topological features in link prediction using supervised learning. The aggregated features, in a social network, are some aggregation functions of the attributes of the nodes. Topological features describe the topology or structure of a social network, and its underlying graph. We evaluated the effectiveness of these features by measuring the performance of different supervised machine learning methods. Specifically, we selected five well-known supervised methods including J48 decision tree, multi-layer perceptron (MLP), support vector machine (SVM), logistic regression and Naive Bayes (NB). We measured the performance of these five methods with different sets of features of the DBLP Dataset. Our results indicate that the combination of aggregated and topological features generates the best performance. For evaluation purposes, we used accuracy, area under the ROC curve (AUC) and F-Measure.
Our selected features can be used for the analysis of almost any social network. This is because these features provide the important characteristics of the underlying graph of the social networks. The significance of our work is that the selected features can be very effective in the analysis of big social networks. In such networks we usually deal with big data sets, with millions or billions of instances. Using fewer, but more effective, features can help us for the analysis of big social networks.
Subjects: Social and Information Networks (cs.SI)
Cite as: arXiv:2006.16327 [cs.SI]
  (or arXiv:2006.16327v1 [cs.SI] for this version)

Submission history

From: Mike Raeini [view email]
[v1] Mon, 29 Jun 2020 19:17:24 GMT (318kb,D)

Link back to: arXiv, form interface, contact.