Deep Generative Models for Node Embedding and Neighborhood Prediction in Dynamic Graphs of Recommendation Systems

In this article, we develop generative models that generate embeddings for graph nodes while using only their initial features without any knowledge about their neighborhoods and connections. Accordingly, we start by generating reference embeddings using a graph neural network (GNN) trained on full graph knowledge. Afterward, we train the generative models, specifically an autoencoder and a generative adversarial network (GAN), which use only the initial node features to generate close and almost indistinguishable embeddings to those generated by the GNN. To this end, we use a customized loss function acting as a strong regularization for our models. It compels them to generate only embeddings with small error values from those generated by the fully fledged model. Using real-world graph datasets, we evaluate the quality of the generated embeddings for different similarity metrics such as the mean-squared error (MSE) and cosine similarity. We also assess their ability in reconstructing an initial graph and predicting the neighborhood of each newly added node. Results show the superiority of the proposed generative models over the conventional ones and that the proposed GAN model outperforms the proposed autoencoder with an efficiency in graph reconstruction exceeding 85% for different datasets.