A Journey Towards Flexible Graph Machine Learning (FlexGML)
By Aref Einizade

Aref Einizade will give a talk on “A Journey Towards Flexible Graph Machine Learning (FlexGML)”

Abstract

Due to the great potential of modeling and analyzing the multi-sensor data as instances of graph-structured data, the talk discusses the past and ongoing speaker’s research in extending data science from regular Euclidean structures to flexible non-regular ones, particularly through graph data science (GDS) and graph machine learning (GML). This mainly includes graph neural networks (GNNs) for learning from graph-structured data and graph signal processing (GSP), which generalizes classical signal processing for diverse applications like neuroscience and spatiotemporal forecasting. Precisely, the talk identifies key challenges in current GML approaches as: 1) The reliance on accurate knowledge of the underlying graph structure, which can be problematic in noisy or complex systems like brain connectivity; 2) Limitations in handling multiple interacting graphs, which are needed in applications like video processing and brain data analysis; 3) Underdeveloped methods for continuous graph filters, such as the lack of frameworks for defining partial differential equations (PDEs) on multi-domain graphs; 4) GNNs face issues like over-smoothing and over-squashing, hindering long-range message passing; 5) Many GDS methods assume undirected graphs, while real-world systems (e.g., traffic modeling) might involve directed graphs as well. The speaker’s research focuses on addressing these limitations by approaching towards a more flexible GML framework called FlexGML with a broad range of real-world applications from neuroscience and biomedical signal and image processing to spatiotemporal traffic and weather forecasting.

Biography

Since November 2023, Aref EINIZADE has been a postdoctoral researcher at the Multi-Media (MM) team, Télécom Paris, Institut Polytechnique de Paris, working on the intersection of graph machine learning (GML), graph signal processing (GSP) and graph neural networks (GNNs) to alleviate the fundamental limitations of the current GML algorithms and devise novel and theoretically supported frameworks mainly when facing with multiple multi-modal graphs. Prior to that, he got his Ph.D. in February 2023 in Electrical Engineering from the Sharif University of Technology, Tehran, Iran, by developing GSP and GML algorithms mainly in the challenging cases of unknown graphs and their applications in biomedical (mostly the brain) data processing. His interest lies within analytical machine learning (ML) and signal processing (SP), in which graph-centric concepts of classic ML and SP are theoretically studied in a rigorous manner and their real-world applications are vastly explored from classic statistical concepts (e.g., Blind Source Separation of Graph Signals) to modern topics (like Continuous GNNs). Now, his focus is mostly on more generalized structures like Simplicial Complexes and Hypergraphs to extend the theoretical and experimental results in more complex structures.