Author

Drew Ligman

Date of Award

6-10-2021

Document Type

Thesis

Publisher

Santa Clara : Santa Clara University, 2021.

Department

Computer Science and Engineering

First Advisor

Zhiqiang Tao

Abstract

With the increasing size and complexity of machine learning datasets, obtaining highly performing prediction models in various tasks has become increasingly difficult. In particular, the processs of hyperparameter optimization (HPO) contributes a significant portion of this cost. This work examines a specific graph-machine learning model, graph convolutional networks (GCN), to derive a hyperparameter configuration with optimal performance across a variety of datasets. We motivate our configuration theoretically and validate it empirically through comprehensive experimentation. We find that for GCN semi-supervised classification tasks, our configuration performs nearly optimally when compared against traditional HPO while only requiring a fraction of the budget. We further propose using this configuration to warm-start subsequent HPO as a means of accelerating its convergence.

Share

COinS