DeHIN: A Decentralized Framework for Embedding Large-scale Heterogeneous Information Networks

• DOI: 10.48550/arxiv.2201.02757
• Publication Date: 2022-01-01
• Authors: Mubashir Imran; Hongzhi Yin; Tong Chen; Zi Huang; Kai Zheng
• Keywords: Subnetwork

Modeling heterogeneity by extraction and exploitation of high-order information from heterogeneous networks (HINs) has been attracting immense research attention in recent times. Such network embedding (HNE) methods effectively harness the small-scale HINs. However, real world, size HINs grow exponentially with continuous introduction new nodes different types links, making it a billion-scale network. Learning node embeddings on such creates performance bottleneck for existing HNE that are commonly centralized, i.e., complete data model both single machine. To address large-scale tasks strong efficiency effectiveness guarantee, we present \textit{Decentralized Embedding Framework Heterogeneous Information Network} (DeHIN) this paper. In DeHIN, generate distributed parallel pipeline utilizes hypergraphs order to infuse parallelization into task. DeHIN presents context preserving partition mechanism innovatively formulates large HIN as hypergraph, whose hyperedges connect semantically similar nodes. Our framework then adopts decentralized strategy efficiently adopting tree-like pipeline. Then, each resulting subnetwork is assigned worker, which employs deep maximization theorem locally learn receives. We further devise novel alignment scheme precisely project independently learned all subnetworks onto common vector space, thus allowing downstream like link prediction classification.