Graph-convolutional neural networks are growing increasingly popular, but most of them limit themselves to simply considering user-item interactions, even though additional information is often available in datasets such as context and side-information.
In this paper, we present two ways to incorporate side-information and contextual information into the prediction model of a graph-convolutional neural network named CSGCN-IS and CSGCN-ADJ.
Including this additional information allows us to not only improve density of the graph structure, but also to generate recommendations for a specific context that the user is currently in.
We empirically evaluate the models in both a context-specific setting as well as a non-context-specific on four different real-world datasets, comparing with several relevant GCN and FM models.
The non-context specific evaluation employs an 80-20% training and test data split, and shows improvements in performance from 0.07%-10.01%, as well as a decrease on certain datasets of up to 5.09%.
The context-specific evaluation shows both significant improvements and decreases.
An ablation study is also conducted, showing that the inclusion of context and side-information for CSGCN-ADJ does little to improve performance for the non-context specific setting.
For the context-specific setting, the ablation study shows that the performance of CSGCN-IS increases when context and side-information are included, whereas CSGCN-ADJ sees little difference.