Composition of Heterogeneous Node Embeddings - Unlocking the Power of Heterogeneous Graph Representation

Heterogeneous graphs have high representation power, which can be maximized through node embeddings. Important embedding approaches are based on node features and node metapaths, applied individually. This paper proposes the creation of heterogeneous composition node embeddings, which are based on local node features, features from node neighbors, and node metapaths. This results in two types of composition embeddings: Features + Metapaths and Aggregated + Metapaths. Experiments have demonstrated superior performance compared to the baseline. In the experiments, our composition Aggregated Features + Metapaths embedding achieved a Micro-F1 score of 65.89% compared to 61.53% from the baseline, highlighting its effectiveness. Additionally, this paper also evaluates alternative models with these embedding compositions that outperform the state-of-the-art approach.