Graph Embedding with Personalized Context Distribution.

Graph representation learning embeds graph nodes in a low-dimensional latent space, which allows for mathematical operations on nodes using low-dimensional vectors for downstream tasks, such as link prediction, node classification, and recommendation. Traditional graph embedding methods rely on hyper-parameters to capture the rich variation hidden in the structure of real-world graphs. In many applications, it may not be computationally feasible to search for optimal hyper-parameters. In this work, built on WatchYourStep which a graph embedding method leveraging graph attention, we propose a method that utilizes node-personalized context attention to capture the local variation in a graph structure. Specifically, we replace the shared context distribution among nodes with learnable personalized context distribution for each node. We evaluate our model on seven real-world graphs and show that our method outperforms the state-of-the-art baselines on both link prediction and node classification tasks. We further analyze the learned node context distribution to provide insights into its connection to graph structural properties.