The paper addresses the task of learning whole-graph representations for signed networks, which is an important but underexplored problem. The authors make the following key contributions:
Methodological contribution: They propose two approaches for learning whole-graph representations of signed networks - SG2V (a signed generalization of Graph2vec) and WSGCN (a whole-graph generalization of the signed vertex embedding method SGCN). They define several variants of these methods.
Resource contribution: They constitute and share a benchmark dataset of signed graphs annotated for classification, comprising three distinct collections. This is the first such benchmark for signed graph classification.
Experimental contribution: They evaluate their proposed methods on the benchmark and show that they outperform a baseline approach. The SG2V and WSGCN methods achieve F-measure scores of 73.01 and 81.20 respectively, compared to 58.57 for the baseline.
The paper first provides background on signed graphs and existing graph representation learning methods. It then describes the three signed graph datasets used for evaluation. The core of the paper details the proposed SG2V and WSGCN methods, which are adaptations of unsigned whole-graph embedding and signed vertex embedding approaches to handle signed graphs at the whole-graph level. The experimental results demonstrate the effectiveness of the new signed whole-graph representation learning methods compared to the baseline.
To Another Language
from source content
arxiv.org
Deeper Inquiries