Combining Relations and Text in Scientific Network Clustering

Abstract

In this paper, we present different combined clustering methods and we evaluate their performances and their results on a dataset with ground truth. This dataset, built from several sources, contains a scientific social network in which textual data is associated to each vertex and the classes are known. Indeed, while the clustering task is widely studied both in graph clustering and in non supervised learning, combined clustering which exploits simultaneously the relationships between the vertices and attributes describing them, is quite new. We argue that, depending on the kind of data we have and the type of results we want, the choice of the clustering method is important and we present some concrete examples for underlining this.

I-Louvain: An Attributed Graph Clustering Method

“…The experimental study of section 5 confirms that clustering, based on the relational information and attributes provides more meaningful clusters than methods taking into account one type of data (attributes or edges) or than ToTeM which exploits attributes and edges [6].…”

“…In the following experiments, we study the robustness of our method to various degradations of an artificial network and we compare its performances, according to the accuracy as well as the normalized mutual information, with K-means, Louvain and ToTeM. Among the methods exploiting the both kinds of data (relationships and attributes), Totem has been retained because it has been showned experimentally that it provides better results than simpler methods [6,5] Finally, the last experiments aim at studying the impact of increasing the number of vertices and edges on the run-time evolution.…”

“…Our first experiments aim at evaluating on a real dataset the performances of I-Louvain, which exploits attributes and relational data, compared with methods based only on one type of data, K-means for the attributes and Louvain for the relations and with ToTeM, an other community detection method designed for attributed graphs which exploits the two types of information, notably numerical attributes [6]. In the following experiments, we study the robustness of our method to various degradations of an artificial network and we compare its performances, according to the accuracy as well as the normalized mutual information, with K-means, Louvain and ToTeM.…”

“…Recently, some of these methods have been compared and these experiments have confirmed that the detection of communities in an attributed graph is not a trivial problem [6,11]. To solve it efficiently, we consider that the attributes and the relational information must be exploited simultaneously and this is not the case for several methods cited.…”

“…Generally, this is not the case in clustering of vertices where only the relationships between the vertices are used, nor in unsupervised classification based only on the attributes. Recently, several methods have been proposed to take into account the relational information as well as the attributes in the aim to detect patterns in attributed graphs [26,31] or to tackle this problem of hybrid clustering [6,11]. In this article, we propose a method, called I-Louvain, which allows to partition the vertices of an attributed graph when numerical attributes are associated to the vertices.…”