Çizge sinir ağlarında temsil öğrenimini geliştirmek

Real-world data often comprises interconnected entities rather than isolated objects. Traditional neural networks struggle to process irregular structures inherent in prominent domains such as social networks, bioinformatics, and physical systems that have extensively benefited from deep learning. Graph neural networks (GNNs) following the message-passing paradigm have emerged as a versatile and effective framework for processing relational data with arbitrary structural characteristics. However, the efficiency of GNNs comes with challenges in representation learning that impact downstream task performance. This thesis addresses two major challenges occurring in deep GNNs: over-smoothing and over-squashing. The message-passing mechanism, by design, is prone to producing almost identical representations after a certain number of steps, causing an over-smoothing effect. We propose an adaptive channel-wise message-passing approach to alleviate over-smoothing. The proposed model, Channel-Attentive GNN (CHAT-GNN), learns how to attend to neighboring nodes and their feature channels. Thus, more diverse information can be transferred between nodes during message-passing. Since information between distant nodes is transmitted through sequential message-passing between adjacent nodes, community-connecting nodes must compress substantial information into low-dimensional feature spaces, resulting in the over-squashing effect. We propose incorporating local virtual nodes that increase the feature capacity of central nodes and the number of paths through central regions. Moreover, our shared virtual node embedding approach enables communication with a range greater than the receptive field of GNNs. We compare both approaches with state-of-the-art baselines in terms of downstream task performance and conduct detailed experiments to validate the effectiveness of CHAT-GNN and local virtual nodes.