Item

Abstract

In machine learning frequently the data appear in the form of graphs. Biological systems modelling, image processing, and social networks, are instances of applications that require graph learning. Geometric Deep Learning is an arising field of techniques that has extended deep neural applications to non-Euclidean domains. Graph Convolutional Networks have created a promising pathway in graph-based tasks and have achieved advanced performance on semi-supervised and supervised learning. However, these methods have challenged the computational demand of high dimensional data and attempted to train neural networks in an analogous scale. High-dimensional data are a worthy opponent of Artificial Intelligence and machine learning as their high dimensional nature affects the performance of optimization and classification. Dimensionality reduction algorithms aim to support machine learning by embedding the data to the low dimensional space. Uniform Manifold Approximation and Projection (UMAP) is a novel dimensionality reduction technique that is gaining traction in the area of visualization a n d pre-processing in machine learning. In this study, we have applied dimensionality reduction algorithms with main focus on UMAP to accomplish disease prediction using the proposed Graph Convolution Network of Parisot et al. (2017). The model is trained o n semi-supervised learning to achieve node classification on a population graph. The performance of the model is being explored using two different datasets and a comparison in classification with a ridge classifier is implemented. The data concern individuals with Attention Deficit Hyperactivity Disorder (ADHD) and patients with different forms of frontotemporal lobar degeneration (FTLD). The objective of this dissertation is to apply UMAP and other dimensionality reduction techniques for graph-based classification to improve performance applied on different neurological diseases.