Executive Summary : | Graphs and networks are used in various applications, including molecule discovery and epidemic tracing. Recently, there has been an increasing interest in developing deep learning models for complex network structures, often called graph neural networks (GNNs). These models compress information about the neighborhood structure around each node into a low-dimensional vector called node embedding or node representation vector. However, existing GNNs often suffer from several limitations, such as their invariance to permutation of node-neighbors through a symmetric message passing protocol. This project aims to design deep learning methods for representation learning for complex networks to overcome these challenges. The proposed method would be useful in link prediction, node classification, information diffusion, and graph retrieval. To solve these problems, asymmetric node embeddings using sequential encoders are designed, which characterize dependencies between node neighbors more effectively than symmetric GNNs. However, they are now heavily sensitive to permutations of node neighbors. The problem is addressed by a permutation de-sensitization network driven by the Sinkhorn network, which aims to minimize training objective with respect to trainable parameters. Another challenge is learning similarity between graphs in the context of graph retrieval. A representation learning model is built, which takes two sets of embeddings corresponding to two graphs as input and outputs a similarity score between those two graphs. This problem has applications in graph retrieval, where the goal is to retrieve a set of relevance corpus graphs for a given query graph. |