AttrHIN: Network Representation Learning Method for Heterogeneous Information Network

Abstract

Network representation learning can map complex network to low dimensional vector space, capture the topological properties of network, and reduce the time complexity and space complexity of the algorithm. However, most of the existing network representation learning (NRL) methods are for homogeneous networks, while the real-world networks are usually heterogeneous, therefore, it is more practical to provide an intelligent insight into the evolution of heterogeneous networks. In this paper, we propose a novel heterogeneous network embedding method, called AttrHIN, which adopts weighted metapath-based random walks strategy, and can make full use of the attribute information to capture the latent features. AttrHIN is suitable for the different types of nodes in heterogeneous networks. Extensive experimental results show that compared with the state-of-art algorithms, AttrHIN achieves better results in Macro-F1 and Micro-F1 for multi-class node classification and Link Prediction on several real-world network datasets.

Attributed network embedding based on self-attention mechanism for recommendation method

“…Additionally, they increase model flexibility by decomposing user preferences into different anchor vector representations, making them better suited to adapt to varying user requirements and deliver more personalized recommendations. However, regarding why MGAT outperforms MMGCN, it relates to differences in model structures, a conclusion that aligns with the findings in Reference 16 . In summary, these factors collectively contribute to MGAT and MMGCN surpassing DEEPWALK in terms of recommendation performance and, in some aspects, making MGAT perform even better.…”

“…To address these challenges, the MGAT model 16 first utilizes user-item interaction records and pre-trained modality-specific features to construct a unimodal graph for each modality. Subsequently, it employs Graph Attention Networks (GAT) and Gated Recurrent Units (GRU) to learn high-order neighbor information within the graph and capture local semantic information, further enhancing recommendation performance.…”