Search

Graph Neural Networks (GNNs) have emerged as a powerful class of machine learning techniques capable of processing graph-structured data, showing immense promise in molecular property prediction. This thesis compares the performance of two specific GNNs—Graph Attention Networks (GATs) and Attentive FP models with Graph Convolutional Networks (GCNs) in predicting the biological activities of protein targets across several protein families. Our findings indicate that while GCNs are highly effective for molecular property prediction, GATs and Attentive FP models also offer competitive performance, with GATs showing particular promise for enzymes and transporters. The experiments suggest that the choice of model should be tailored to specific families of target proteins, highlighting the need to consider the particular protein family when selecting GNNs for predictive modeling in drug discovery.

Graph neural network (GNN) is one of the most widely used methods that leverage relational information in the data to learn and make predictions. Fraud detection is a challenging task considering the nature of fraudulent transactions which changes drastically from one case to another as fraudsters often collude to hide their abnormal behavior/features. To this end, GNN has a fitting application because it leverages graph structure to learn relational information to distinguish malicious transactions from legitimate ones. This study implements various GNNs such as graph convolution network (GCN), graph attention network (GAT), and modified GAT to predict fraudulent Bitcoin transactions. It focuses on benchmarking the results of two versions of GAT against GCN to demonstrate the superior predictive power of the attention mechanism. The two versions include conventional GAT and modified GAT, the latter consists of a dynamic attention mechanism. The two versions of the GAT model are also compared in detail. GNN has been used to detect fraud or anomalies for various practical implementations such as financial transactions, credit cards, and customer reviews. However, a detailed study focusing on the two versions of GAT and benchmarking it against GCN has not yet been conducted. We show that GAT has an enhanced ability to predict fraudulent transactions. The excellent predictive performance of GAT gives a clear indication that it could play a vital role in detecting broader cryptocurrency fraud. Finally, this study discusses the challenges of building an explainable GNN models.

Hyperspectral image classification can be a difficult task due to a lack of ground-truth labels and spectral variability. For hyperspectral images (HSIs) of historical artifacts, which are generally man-made objects, this is generally a significant challenge. Aiming to differentiate subtle spectral differences between varied pigments, we introduce graph- based learning networks and a powerful learning pipeline for realizing automatic pigment mapping on the maps. First, we develop a Graph Modularity clustering method for pro- ducing trustworthy ground-truth map for pigment classification task. We analyze the influence of graph structure on the Graph Neural Networks and propose Spatial-Spectral Graph Convolutional Network (SSGCN) and Spatial-Spectral Graph Attention Network (SSGAT). By incorporating spatial features and aggregating node feature in terms of spec- tral affinity and spatial context, SSGCN and SSGAT outperform conventional GCN and GAT, and can be directly applied to another unknown region for automatic pigment map- ping without ground-truth labels required. Under a semi-supervised node classification framework, SSGCN and SSGAT achieve high classification accuracy with very limited labeled samples. Compared to a 3D-CNN architecture, our graph-based learning mod- els provide great scalability in model extension and inductive learning for unseen pixels. We apply these methods to Vis-NIR hyperspectral imagery of the Gough Map of Great Britain and the Selden Map of China, both artifacts in the Bodleian Library, University of Oxford. These maps are of great interest to historians as their creation and modification are generally not well understood. The experiments on multiple Regions of Interest in the two maps demonstrate the strength and advantages of our methods. The proposed graph-based learning frameworks bring a promising prospect for processing large-scale HSI classification/clustering of historical artifacts efficiently. This research will aid historians in pigment analysis and codicological studies of artifacts in a more intelligent manner.