HeteEdgeWalk: A Heterogeneous Edge Memory Random Walk for Heterogeneous Information Network Embedding

Most Heterogeneous Information Network (HIN) embedding methods use meta-paths to guide random walks sample from HIN and perform representation learning in order overcome the bias of traditional that are more biased towards high-order nodes. Their performance depends on suitability generated for current HIN. The definition requires domain expertise, which makes results overly dependent meta-paths. Moreover, it is difficult represent structure complex with a single meta-path. In meta-path guided walk, some heterogeneous structures (e.g., node type(s)) not among types specified by meta-path, making this information ignored. paper, HeteEdgeWalk, solution method does involve meta-paths, proposed. We design dynamically adjusted bidirectional edge-sampling walk strategy. Specifically, edge sampling storage recently selected used better network balanced comprehensive way. Finally, classification clustering experiments performed four real HINs in-depth analysis. show maximum improvement 2% at least 0.6% compared baselines. This demonstrates superiority effectively capture semantic HINs.