Multi-order Proximity Graph Structure Embedding

Wang Zhang, Lei Jiang, Huailiang Peng, Qiong Dai, Xu Bai

Abstract: Graph embedding methods convert the flexible graph structure into low-dimensional representations while maintaining the graph structure information. Most existing methods focus on learning low- or high-order graph information, and cause loss of information during the embedding process. We instead propose a new method that can learn low and high order graph information simultaneously. The method fuses structure-preserving model with random walk sampling, which learns multi-order graph structure information more efficiently. Our method also utilizes distance-based weighted negative samples to improve the representations learning. The experimental results indicate that our proposed method provides very competitive results on the node classification, node clustering and graph reconstruction tasks for four benchmark datasets, BlogCatalog, PPI, Wikipedia and email-Eu-core.