TACKLING OUT-OF-DISTRIBUTATION DATA IN GRAPH NEURAL NETWORKS

Content Files

SP23-ECE499-Thesis-Hu, Chuxuan.pdf

Permalink

https://hdl.handle.net/2142/124848

Title: TACKLING OUT-OF-DISTRIBUTATION DATA IN GRAPH NEURAL NETWORKS
Author(s): Hu, Chuxuan
Issue Date: 2023-05-01
Keyword(s): Graph Neural Networks; Out-Of-Distribution Data
Date of Ingest: 2024-10-14T10:47:56-05:00
Abstract: Out-of-distribution (OOD) data has severely affected the performance of graph neural networks (GNNs). For example, when a GNN has been trained to classify citation networks based on the data from a specific time period, its performance drops drastically when switching to other time periods due to distribution drifts. For my senior thesis, I have put forward the first methodology to tackle the problem by co-training multiple GNNs with various training techniques and aligning the resulted encodings under the supervision of Professor Bo Li. We are also the first to rigorously demonstrate that the distribution shifts of graphs come from both node feature shifts and structural shifts. The superiority of my proposed framework has been well verified through comprehensive empirical analysis on the OGB-Arxiv dataset [1], an open-source citation network.
Type of Resource: text
Genre of Resource: dissertation/thesis
Language: eng

The best of ECE undergraduate research