Inductive link prediction in graphs

• Main Author: Wang, Yu Zhen
• Other Authors: Lihui Chen
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2022
• Subjects: Engineering::Computer science and engineering::Computing methodologies::Artificial intelligence; Engineering::Electrical and electronic engineering
• Online Access: https://hdl.handle.net/10356/157521
• Institution: Nanyang Technological University

Predicting the link between a pair of nodes in a graph is an important task for graph data analytics. Most existing research focuses on transductive link prediction, where both nodes already exist in graphs, or their models built are inherently transductive. However, many real-world applications require models to embed unseen nodes, with only their attribute information, into the graph inductively. One of the recent attempts at inductive link prediction proposes a method called Edgeless-GNN, which leverages the power of graph neural networks (GNN) models such as GraphSAGE, by replacing its original computation graph with k-nearest neighbour graph to empower its inductivity. Current application of this state-of-the-art model is more towards citation networks, has not been explored on co-purchase graphs, which can be exploited to resolve recommendation problems. This project analyses the original Edgeless-GNN model, designs and implements Edgeless-GNN with GraphSAGE, also called Edgeless-SAGE, and tested it on Amazon co-purchase networks. Model finetuning and performance evaluations are also conducted, aiming to improve the embeddings for unseen, edgeless nodes into the vector space hence to achieve better performance for inductive link prediction tasks.