Informed Heterogeneous Attention Networks for Metapath Based Learning

Metapath based processing is popular for heterogeneous graphs as it can be focused on relevant relations. However, recent work has shown that important information about the intermediate nodes that form a metapath is lost in the process. This puts it at a disadvantage compared to homogeneous graph processing, where all node information is available. We propose a novel attention mechanism that can be used to incorporate structural intermediate node information into metapath based attention. Combined with a more efficient propagation and aggregation strategy, this improves the performance of metapath based processing on heterogeneous graphs. These adaptations allow us to surpass state-of-the-art performance in node classification tasks from the heterogeneous graph bench-marking suite by up to 2%. We further improve upon the original Heterogeneous Attention Network by up to 8%. The used codebase as well as code for all experimental setups with results are available at https://github.com/wendli01/info_han.