BRAIN CONNECTIVITY NETWORK PREDICTION USING DEEP LEARNING FOR GRAPH TRANSFORMATION

The recent surge of research into Graph Neural Networks (GNNs) has come a long way to approach problems which require structured predictions for “translating” an input graph into a corresponding output graph. Although successful in many applications, conventional GNNs are only able to deal with node translation problems to predict the node attributes or node category of the target graph. In many practical scenarios, node-edge co-transformation is required where both nodes and edges can change during translation process. Characterizing the underlying mechanism of graph topological evolution from a source graph to a target graph has been also addressed by spectral based approaches which are built on the spectral evolution model. The growth of large networks is analyzed by studying the changes in the spectral characteristics of the graph. Motivated by the evolution and transformation of brain connectivity network, my research goal is to develop an end-to end graph translation framework that can optimally handle the change in graph topology (novel technique), and apply that to the real-world domain specific task of brain connectivity network translation problem, such as structural to functional connectivity translation. Despite the rapidly-growing research on graph translation topic, in many applications, these are generally focused on either linear models or computational models that rely heavily on heuristics and simple assumptions. However, depending on the domain, while the topology changes, the relationship could be highly-nonlinear, complex, and contain considerable randomness. Beyond merely doing transformation and prediction, it is also interesting and important to figure out which subgraphs in the input graph majorly influence which subgraphs in the output. Interpretable models that can probe the data automatically and find candidate pairs of input and output subgraphs with strong correlation, are in urgent demand. Additionally, research on GNN explainability on how to generate explanations and importantly how to adjust the model to generate more accurate explanations is becoming more critical. Unlike local explanation models which explain the model prediction per input sample, global explanation techniques aim at providing general insights and high-level understanding of the predictions of a deep graph model. Specifically, they investigate what input graph patterns can lead to a certain GNN behavior or maximize the predicted probability for a certain class and use such graph patterns to explain the class. This is essential in many real-world critical applications and can substantially increase human trust in GNNs’ prediction ability. To address these challenges, my research plan is mainly 3-fold: 1) develop a novel framework for graph translation which can model the change in graph topology, and is capable of learning the stochasticity; 2) propose new post-hoc explainer of our framework that can identify which subgraphs in input strongly influence which subgraphs in output and apply that to the structural to functional connectivity mapping problem which is built upon our proposed framework; 3) design a global GNN Explanation Supervision (GNES) framework that can improve the reasonability of the generated explanations in a global manner, while still keep or even improve the backbone GNNs model performance.

Author

• Etemadyrad, Negar

Identifier

hdl:1920/14381

OAI identifier oai:identifier

oai:MARS:1920/14381