Graph Regression Based on Graph Autoencoders.

We offer in this paper a trial of encoding graph data as means of efficient prediction in a parallel setup. The first step converts graph data into feature vectors through a Graph Autoencoder (G-AE). Then, derived vectors are used to perform a prediction task using both a Neural Network (NN) and a regressor separately. Results for graph property prediction of both models compared to one another and baselined against a classical graph regression technique i.e. Nearest Neighbours, showed that using embeddings for model fitting has a significantly lower computational cost while giving valid predictions. Moreover, the Neural Network fitting technique outperforms both the regression and Nearest Neighbours methods in terms of accuracy. Hence, it can be concluded that using a non-linear fitting architecture may be suitable for tasks similar to representing molecular compounds and predicting their energies, as results signify the G-AE's ability to properly embed each graph's features in the latent vector. This could be of particular interest when it comes to representing graph features for model training while reducing the computational cost.