Hi, I am Mohit (मोहित)

In the era of big data, graphs have emerged as a natural representation of intricate relationships. However, graph sizes often become unwieldy, leading to storage, computation, and analysis challenges. We propose Universal Graph Coarsening (UGC), a framework equally suitable for homophilic and heterophilic datasets. UGC integrates node attributes and adjacency information, leveraging the dataset's heterophily factor. Results demonstrate that UGC preserves spectral similarity while coarsening. In comparison to existing methods, UGC is 4× to 15× faster, has lower eigen-error, and yields superior performance on downstream processing tasks even at 70% coarsening ratios.^¹

Graph structure learning is a core problem in graph-based machine learning, essential for uncovering latent relationships and ensuring model interpretability. However, most existing approaches are ill-suited for large-scale and dynamically evolving graphs. We propose GraphFLEx—a unified and scalable framework for Graph Structure Learning in Large and Expanding Graphs. GraphFLEx mitigates scalability bottlenecks by restricting edge formation to structurally relevant subsets of nodes identified through clustering and coarsening techniques. The framework supports 48 flexible configurations and achieves state-of-the-art performance with significantly improved scalability across 26 diverse datasets.

Graph Coarsening (GC) is a prominent graph reduction technique that compresses large graphs to enable efficient learning and inference. However, existing GC methods generate only one coarsened graph per run and must recompute from scratch for each new coarsening ratio. We introduce a novel framework that combines Locality-Sensitive Hashing (LSH) with Consistent Hashing to enable adaptive graph coarsening. For heterogeneous graphs, we propose a type-isolated coarsening strategy that ensures semantic consistency. Our approach is the first unified framework to support both adaptive and heterogeneous coarsening, achieving superior scalability across 23 real-world datasets.

The emergence of single-cell technologies has revolutionized the study of cellular heterogeneity, generating vast datasets rich in biological insights. However, managing large-scale graph representations of single-cell data remains computationally challenging. We propose a novel approach that utilizes our established coarsening algorithm, integrating locality-sensitive hashing (LSH) to expedite processing without compromising critical data features. This method directly extracts informative, low-dimensional cell representations from raw single-cell RNA sequencing and mass cytometry data, significantly improving processing speed while preserving essential data features. Our method's efficiency, precision, and adaptability represent significant advancements in large-scale single-cell analysis.

Low-dimensional embeddings are central to analyzing and visualizing high-dimensional data. However, widely adopted NE methods assume centralized access to all data, an unrealistic constraint in privacy-sensitive, decentralized environments. We propose SENSE, a geometry-aware, privacy-preserving framework for global neighbor embedding without raw data exchange. SENSE reconstructs global structure using local distance measurements and structured matrix completion, enabling embeddings that preserve both local and global geometry in Euclidean and hyperbolic spaces. It further integrates contrastive learning by deriving cross-client positive and negative pairs from estimated similarities, effectively generalizing negative sampling under structural constraints. Experiments across diverse real-world datasets show that SENSE achieves embedding quality on par with centralized baselines, while offering strong privacy guarantees. Theoretical analysis provides formal bounds on reconstruction fidelity and privacy, establishing conditions under which structure and confidentiality are jointly preserved.

Federated graph learning has emerged as a significant approach for training graph machine learning models across various domains. However, graphs often exhibit non-IID properties in both structure and node features. We propose pFedGNN, a personalized federated graph-based learning approach that utilizes inferred collaboration graphs. The core idea is to model collaboration benefits between clients by inferring a collaboration graph based on pairwise GNN model similarity and graph dataset size. This approach gives significant improvement on GNN node classification as it encourages collaboration primarily among similar and beneficial clients.