Date of Award
2024
Document Type
Thesis (Master's)
Department or Program
Computer Science
First Advisor
Soroush Vosoughi
Second Advisor
Yujun Yan
Third Advisor
Yu-Wing Tai
Abstract
Graph Neural Networks (GNNs) have emerged as powerful tools for modeling and analyzing graph-structured data, gaining prominence in applications such as social network analysis, recommendation systems, and bioinformatics. Despite their success, the opaque nature of GNNs has raised concerns regarding their interpretability and trustworthiness. This thesis addresses the challenge of explainability in GNNs by proposing novel methodologies that enhance the understanding and performance of these models through task-specific explainability mechanisms.
The research introduces techniques that leverage both functional and structural similarity of neurons within GNNs to provide comprehensive explanations. The func- tional similarity is assessed by analyzing gradients, activations, and covariance of neuron responses. Methods such as Neuron Conductance and Integrated Gradients are employed to identify critical neurons and their contributions to the model’s per- formance. Structural similarity is measured by examining the impact of individual neurons on the predictions of specific subgraphs within the GNN. Subgraphs are in- duced by nodes whose predictions change significantly when a neuron is deactivated. These subgraphs are then used to train a contrastive learning model that generates meaningful embeddings, facilitating the clustering of neurons based on their structural roles within the network.
The proposed approach shifts the focus from local explanations, which clarify individual predictions, to global explanations that uncover the overarching principles governing GNN behavior. This holistic perspective allows for the fine-tuning of GNN components to enhance performance on specific tasks. Experimental eval- uations demonstrate the effectiveness of the proposed methods, showing significant improvements in model interpretability and task-specific performance.
Recommended Citation
Cavdaroglu, Barkin, "TASK-SPECIFIC EXPLAINABILITY OF GRAPH NEURAL NETWORKS TO IMPROVE MODEL PERFORMANCE USING FUNCTIONAL AND STRUCTURAL SIMILARITY OF NEURONS" (2024). Dartmouth College Master’s Theses. 130.
https://digitalcommons.dartmouth.edu/masters_theses/130