My name is Adarsh Jamadandi. I have a Master’s degree in Computer Science from Saarland University. My research interests include Graph Representation Learning and Geometric Deep Learning.

I am currently working as a research assistant at the SprintML Lab, studying the generalization behavior of graph neural networks under the supervision of Franziska Boenisch and Adam Dziedzic.

I finished my master’s thesis at the Relational Machine Learning Lab under the supervision of Dr. Rebekka Burkholz.

Previously, I completed my Bachelors in Electronics and Communication Engineering with Dr. Uma Mudenagudi as my advisor.

Updates

May , 2025 Now pre-print alert! Finding Memo(rization) in Graph Neural Networks. Do GNNs also memorize training data? How do we explain the emergence of memorization in GNNs? Why do only certain nodes get memorized? We introduce the first framework for studying label memorization in GNNs, showing that GNNs' implicit bias toward leveraging graph structure in low-homophiliic settings results in a model that memorizes. We also explain why only certain nodes are prone to memorization via a novel label disagreement score.
January , 2025 Now accepted at ICLR 2025! GNNs Getting ComFy : Community and Feature Similarity Guided Rewiring . Why does graph rewiring via spectral gap maximization work? Should we always maximize the spectral gap? What dictates the success of these methods? Would additional information such as node features and latent community structure help better rewire the graph? These are some of the questions we investigate in this work.
September , 2024 Accepted at NeurIPS 2024! Spectral Graph Pruning Against Over-Squashing and Over-Smoothing. We introduce the Braess Paradox for the first time in context of GNNs. We propose a novel graph pruning strategy that can mitigate both over-squashing and over-smoothing and bonus also find graph lottery tickets!
August, 2024 My master's thesis is now online! On the Importance of Graph-Task Alignment for Graph Neural Networks.
December, 2020 Graph of Thrones: Adversarial Perturbations dismantle Aristocracy in Graphs is accepted at AAAI'2021 Student Poster Program, and the extended version is accepted at DiffGeo4DL, NeurIPS 2020.
November, 2020 Probabilistic Word Embeddings in Kinematic Space, accepted at ICPR, 2020!