Representation Learning

New techniques to learn effective node/graph representations

Graphs are a highly informative, flexible, and a natural way to represent data. Machine Learning for graphs (MLG), whereby classical ML is generalized to irregular graph domains, has enjoyed a recent renaissance, leading to a dizzying array of models and their applications in drug design, Web and recommender systems, social networks, and many more. The key challenge in MLG is to effectively learn to encode discrete graph structure into low-dimensional vector representations or embeddings which are then used in various prediction tasks.

The most popular GRL approaches include (i) random-walk based methods which first transforms the graph into node sequences, in which, the co-occurrence of two nodes measures the structural distance between them; (ii) matrix-factorization based methods which uses low ranked decomposition of some node similarity matrix as node embeddings; and (iii) graph neural networks (GNNs) which encode graph structure directly in the latent representations via neighborhood aggregation strategies. While these methods have shown state of the art improvements in various node/edge/graph level tasks on graphs, there are still open questions about their effectiveness for datasets with varying structural and task properties. On the one hand, we develop new datasets and benchmarks to audit the learning capabilities of existing methods. On the other hand we focus on developing new methods specifically for graphs with certain structural properties and under certain task conditions.