Encoding feature set information in heterogeneous graph neural networks for game provenance

Game Provenance has been proposed and employed for Game Analytics tasks as they capture game session data in detail and allow exploratory analysis and visualizations. Games are highly heterogeneous models with several interacting agents and game-world environment elements. Game Provenance Graphs can accommodate the heterogeneous nature of such applications with different types of nodes and edges that tend to share information across themselves, enhancing cause-effect features rarely addressed by any other approach. On the other hand, existing Heterogeneous Graph Neural Network (HGNN) solutions disregard node feature information, overlooking shared features across distinct node types, and rely on naive approaches, such as projecting each type of node to the same n-dimensional space. We conjecture that leveraging heterogeneous feature information is essential for tackling Game Analytics tasks, especially through Machine Learning based models. To achieve that, we propose a novel approach that allows HGNNs to leverage Game Provenance Graphs' heterogeneous node feature information. Hence, we introduce in this paper three strategies for Heterogeneous Graph Representation Learning that encodes feature set information into the HGNN architecture and projects feature values leveraging similarities across such feature sets. We conduct experiments on two Game Provenance Graphs datasets, the Smoke Squadron and the Game Provenance Profile datasets, which gather game session data from different games. Our results show that encoding feature set information in the representation learning process improves the outcomes of GNN models in non-disjoint feature datasets.