PLSGAN: A Power-Law-modified Sequential Generative Adversarial Network for Graph Generation
| Autor: | Haiwei Zhang, Yanting Yin, Yining Lian, Qijie Bai, Xiaojie Yuan |
|---|---|
| Rok vydání: | 2020 |
| Předmět: |
Theoretical computer science
business.industry Computer science Deep learning 02 engineering and technology 010501 environmental sciences Random walk 01 natural sciences Power law symbols.namesake Adversarial system 020204 information systems 0202 electrical engineering electronic engineering information engineering symbols Topological graph theory The Internet Pareto distribution Artificial intelligence business Generative grammar 0105 earth and related environmental sciences |
| Zdroj: | Web Information Systems Engineering – WISE 2020 ISBN: 9783030620042 WISE (1) |
| DOI: | 10.1007/978-3-030-62005-9_9 |
| Popis: | Characterizing and generating graphs is essential for modeling Internet and social network, constructing knowledge graph and discovering new chemical compound molecule. However, the non-unique and high-dimensional nature of graphs, as well as complex community structures and global node/edge-dependent relationships, prevent graphs from being generated directly from observed graphs. As a well-known deep learning framework, generative adversarial networks (GANs) provide a feasible way, and have been applied to graph generation. In this paper, we propose PLSGAN, a Power-Law-modified Sequential Generative Adversarial Network to address aforementioned challenges of graph generation. First, PLSGAN coverts graph topology to node-edge sequences by sampling based on biased random walk. Fake sequences are produced by trained generator and assembled to get target graph. Second, power law distribution of node degree is taken into consideration to modify the learning procedure of GANs. Last, PLSGAN is evaluated on various datasets. Experimental results show that PLSGAN can generate graphs with topological features of observed graphs, exhibit strong generalization properties and outperform state-of-the-art methods. |
| Databáze: | OpenAIRE |
| Externí odkaz: |
|