Learning from Networks: Graph AI for Real-World Data

Supervisor: Haleh Dizaji, Sareh Maleki Author: Phillip Lutz

Abstract

In the Edge AI team, we develop and evaluate machine learning (ML) methods for graph-structured data across domains such as knowledge graphs, bioinformatics, and social networks, and apply them in distributed and federated environments. We investigate graph learning approaches, including graph neural networks (GNNs) and embedding-based methods, to understand their performance across diverse application scenarios, including resource-constrained edge devices. The research includes the systematic evaluation of these models in terms of predictive accuracy, computational efficiency, and energy consumption, and provides appropriate models for each scenario. This enables us to analyze trade-offs between model complexity and accuracy and to identify suitable approaches for specific use cases. In addition, we explore the application of distributed and federated learning techniques for graph data, aiming to enable scalable and privacy-preserving learning in such environments.

During this internship, the intern will gain hands-on experience with graph-based ML methods and apply them to a real-world problem. After an initial introduction, the intern will preprocess the ABIDE dataset (preprocessed neuroimaging data), including extracting and embedding useful features such as time-series data and representing it as a graph. Subsequently, the intern will implement and apply multiple graph learning models (such as GCN, GAT, and APPNP) to perform a graph classification task (autism detection). Finally, the intern will evaluate and compare the predictive accuracy of different models and provide a summary of the results.