Which type of classifier to use for networked data, connectivity based or feature based?
Files
(Published version)
Date
2018
Authors
Zhang, Z.
Li, J.
Wang, H.
Liu, L.
Liu, J.
Editors
Hacid, H.
Advisors
Journal Title
Journal ISSN
Volume Title
Type:
Conference paper
Citation
Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics), 2018 / Hacid, H. (ed./s), vol.11233 LNCS, pp.364-380
Statement of Responsibility
Conference Name
19th International Conference on Web Information Systems Engineering, WISE 2018 (12 Nov 2018 - 15 Nov 2018 : Dubai, United Arab Emirates)
Abstract
Multi-label classification of social network data has become an important problem. Two types of information have been used to classify nodes in a social network: characteristics of nodes, and the connectivity between nodes. Existing classification methods can be categorized to two types too, feature based methods, and connectivity based methods. We observe that there are no one size fits all classification methods, since the performance is data dependent, but in general node’s class labels are determined by two factors, personal preference and peer influence. However, some data sets are personal preference dominated and are suitable for feature based methods, whereas some data sets are peer influence dominated and are suitable for connectivity based methods. The challenge then is how to judge if a data set is personal preference dominated or peer influence dominated, so a suitable classification method can be selected for its classification. In this paper, we develop a causality based criterion to determine the characteristics of a data set. Experiments on real-world data sets demonstrate the criterion can predict the suitability of a classification method for a data set.
School/Discipline
Dissertation Note
Provenance
Description
Access Status
Rights
Copyright 2018 Springer Nature
Access Condition Notes: Accepted manuscript available after 1 January 2020