Logo image
Multimodality-enhanced graph generation and multimodality-driven graph convolutional networks: a thesis in Data Science
Thesis   Open access

Multimodality-enhanced graph generation and multimodality-driven graph convolutional networks: a thesis in Data Science

Rohan Gonjari
Master of Science (MS), University of Massachusetts Dartmouth
2023
DOI:
https://doi.org/10.62791/20314

Abstract

In the field of multi-view machine learning, understanding and leveraging the inherent structure of multi-view data poses significant challenges. This research presents two novel contributions: Multimodality-Enhanced Graph Generation (MEGG) and Multimodality-Driven GCN (MDGCN). These methodologies are applicable to various multi-view data scenarios, providing flexibility and adaptability. MEGG fuses ”within” and ”between” graphs to generate a comprehensive representation of the data, leveraging complementary information from multiple modalities. The resulting fused graph serves as input to MDGCN, a custom architecture designed for multi-view data analysis. MDGCN comprises graph convolutional and fully connected layers, enabling effective modeling and classification. In addition, this work explores and compares different fusion strategies, presenting a comprehensive examination of their impacts on multi-view GCN methods. This examination includes an assessment of existing graph fusion methods and GCN architectures for multi-view data, offering a contextual understanding of their relative strengths and weaknesses. To evaluate the effectiveness of the methods, this study conducts meticulous training and comprehensive evaluations using two distinct datasets: the MNIST dataset and the Brain Activity Dataset. Comparative analyses against traditional models such as K-Nearest Neighbors (KNN), Support Vector Machines (SVM), Neural Networks (NN), and Decision Trees highlight the superior performance of the proposed methodologies. The flexible nature of MEGG and MDGCN opens up new possibilities for harnessing the inherent structure of multi-view data in various domains.
pdf
Gonjari R. COE MS Thesis 20231.05 MBDownloadView
CC BY-NC-ND V4.0 Open Access

Metrics

122 File views/ downloads
58 Record Views

Details

Logo image