HIN2Grid: A disentangled CNN-based framework for heterogeneous network learning

Recently, graph convolutional networks (GCNs) have been applied to heterogeneous information network (HIN) learning and have shown promising performance. However, the performance of GCNs degrades attributed to the recursive propagation, which leads to an indistinguishable embedding for the distinctly heterogeneous node. Besides, the inherently coupled paradigm of GCNs limits their applications on large-scale graphs. In this paper, we tackle these problems by proposing a disentangled framework named Heterogeneous Information Network to Grid (HIN2Grid) for heterogeneous network learning. We innovatively design an effective and efficient strategy to transform the graph data into semantic-specific grid-like data, which can be effectively processed by convolutional neural networks (CNNs), thus explicitly overcoming the drawbacks of the inherent paradigm of GCNs. Such a CNN-based learning scheme also contributes to extracting more expressive features and consume less time and memory. We further propose dual attention mechanisms to capture the importance of various grid-like data and heterogeneous semantics, thus providing interpretability and robustness for HIN2Grid. We conduct experiments on four datasets and the results show that HIN2Grid significantly outperforms the stateof-the-art methods, gaining a improvement on node classification of about 2% to 10% and a 2 to 5 times promotion on running speed.

Automated summary

This study introduces a disentangled framework called HIN2Grid, which transforms graph data into semantic-specific grid-like data for efficient processing by CNNs, explicitly improving the performance of heterogeneous network learning. Dual attention mechanisms are proposed to enhance interpretability and robustness.