Research Projects

One of the efficient classes of deep learning models is Graph Neural Networks (GNNs) by which graph-structured data is operated; where nodes (vertices) and edges (relationships) are used to store information. GNNs aggregate information through a process from its neighbors to represent each node (or, the graph) which is called message passing. The black-box issue is one of the critical challenges of unknown behavior of GNN to arrive at its predictions due to opaque, complex, fuzzy, and entangled processes. It points out concerns about fairness, trust, debugging, and accountability. This research proposal aims to mitigate the same in Graph Neural Networks (GNNs). Most of the existing explanation methods are post-hoc approximations which are not part of the original learning process. For example, with small input variations or, a small model, there will be a significant change in explanations. Several explainers are expensive computationally and across large graphs or nodes, they do not generalize well. This proposed research will help to overcome this by incorporating fuzzy graph covering and learn able Ordered Weighted Averaging (OWA) operators in the same frame. This is a cutting-edge and excellent combination for potential development. Together, a powerful framework is formed for learning, reasoning, and decision-making in an uncertain environment. Fuzzy graphs are capable of capturing vague, uncertain, and imprecise relationships between entities. To analyze local patterns, the covering of fuzzy graphs decomposes the complex graph for providing interpretable, molecular structures for GNN decisions. With tunable weights, OWA operators allow non-linear, interpretable aggregation which is very helpful to explain the influences of each neighbor in GNN message passing. Development of more robust algorithms associated with Gated Graph Neural Networks (GGNN) and Graph Isomorphism Networks (GIN) is also included as a beneficial part of understanding the computational outcomes and efficiency of the study. This research will encode structure-aware inductive bias like, for different domains, decision-makers can use fuzzy trees or cliques. It is contained with domain knowledge and balanced learning. Additionally, this study helps to possibly reduce over-smoothing by allowing joint learning of feature aggregation and fuzzy graph cover. We will use datasets related to real-world problems like healthcare systems and fuzzy decision-making systems to show the applicability of the developed methodology and check whether it gives reliable results or not.