DHNE: Network Representation Learning Method for Dynamic Heterogeneous Networks

Abstract

Analyzing the rich information behind heterogeneous networks through network representation learning methods is signifcant for many application tasks such as link prediction, node classifcation and similarity research. As the networks evolve over times, the interactions among the nodes in networks make heterogeneous networks exhibit dynamic characteristics. However, almost all the existing heterogeneous network representation learning methods focus on static networks which ignore dynamic characteristics. In this paper, we propose a novel approach DHNE to learn the representations of nodes in dynamic heterogeneous networks. The key idea of our approach is to construct comprehensive historical-current networks based on subgraphs of snapshots in time step to capture both the historical and current information in the dynamic heterogeneous network. And then under the guidance of meta paths, DHNE performs random walks on the constructed historical-current graphs to capture semantic information. After getting the node sequences through random walks, we propose the dynamic heterogeneous skip-gram model to learn the embeddings. Experiments on large-scale real-world networks demonstrate that the embeddings learned by the proposed DHNE model achieve better performances than state-of-the-art methods in various downstream tasks including node classifcation and visualization. INDEX TERMS Dynamic heterogeneous networks, network representation learning, random walk, skip-gram model.

AttrHIN: Network Representation Learning Method for Heterogeneous Information Network

“…The representation learning problem in heterogeneous networks has been briefly introduced in [28]. In specific, we learn from the definition of heterogeneous networks in [29] and claim the learning problem.…”

Dynamic Heterogeneous User Generated Contents-Driven Relation Assessment via Graph Representation Learning

“…where N c (i) denotes the neighborhoods of node i with type c, and θ is the set of parameters. Generally, the transition probability p(m|i; θ) is normalized by the softmax function [18,36,40].…”

“…• DBLP [42] (https://dblp.uni-trier.de) (15 December 2021): The DBLP dataset comprises academic literature information in the field of computer science. In this experiment, we adopt a subset of the DBLP dataset collected by [18], and compress the information into 19 snapshots, which contains 3 types of nodes, i.e., authors, papers and venues.…”

“…• DHNE [18]: A network representation learning method for dynamic heterogeneous networks, which construct historical-current networks snapshots in timelines and capture heterogeneous semantic information under the guidance of metapaths. • DyHATR [19]: DyHATR utilize a hierarchical attention mechanism to learn heterogeneous information and capture network dynamic evolutional patterns via a temporal-attention-based recurrent neural network.…”

“…AMiner [43] (https://www.aminer.cn/data) (15 December 2021): AMiner is a big data mining and service system platform which helps researchers to mine rich academic information. In this paper, we use the evolved dynamic heterogeneous network released by [18], which establishes the relationships among authors, articles and conferences.…”