Name
Abstract | ||
|---|---|---|
Generating general-purpose vector representations of networks allows us to analyze them without the need for extensive feature-engineering. Recent works have shown that the hyperbolic space can naturally represent the structure of networks, and that embedding networks into hyperbolic space is extremely efficient, especially in low dimensions. However, the existing hyperbolic embedding methods apply to static networks and cannot capture the dynamic evolution of the nodes and edges of a temporal network. In this paper, we present an unsupervised framework that uses temporal random walks to obtain training samples with both temporal and structural information to learn hyperbolic embeddings from continuous-time dynamic networks. We also show how the framework extends to attributed and heterogeneous information networks. Through experiments on five publicly available real-world temporal datasets, we show the efficacy of our model in embedding temporal networks in low-dimensional hyperbolic space compared to several other unsupervised baselines. We show that our model obtains state-of-the-art performance in low dimensions, outperforming all baselines, and has competitive performance in higher dimensions, outperforming the baselines in three of the five datasets. Our results show that embedding temporal networks in hyperbolic space is extremely effective when necessitating low dimensions. |
Year | DOI | Venue |
|---|---|---|
2021 | 10.1007/s10618-021-00774-4 | DATA MINING AND KNOWLEDGE DISCOVERY |
Keywords | DocType | Volume |
Hyperbolic embedding, Network embedding, Graph embedding, Representation learning, Temporal networks, Unsupervised learning | Journal | 35 |
Issue | ISSN | Citations |
5 | 1384-5810 | 0 |
PageRank | References | Authors |
0.34 | 0 | 5 |
Name | Order | Citations | PageRank |
|---|---|---|---|
| Lili Wang | 1 | 1 | 2.71 |
| Chenghan Huang | 2 | 1 | 2.03 |
| Weicheng Ma | 3 | 4 | 6.16 |
| Ruibo Liu | 4 | 1 | 5.76 |
| soroush vosoughi | 5 | 50 | 9.78 |