Projects

  • Parallel Sparse Matrix Algorithms
    • Combinatorial BLAS
    • FusedMM: A Unified SDDMM-SpMM Kernel [Github Link][Paper] We developed a fused matrix multiplication kernel that unifies sampled dense-dense matrix multiplication and sparse-dense matrix multiplication under a single operation called FusedMM. It can tune performance using a code generator and perform equally well on Intel, AMD, and ARM processors.
    • SpLib
  • Graph Machine Learning Algorithms
    • Graph Neural Networks
    • Force2Vec: [Github Link][Paper] A parallel method for unsupervised graph embedding. It can generate embedding of a graph having nearly 40 million vertices and on average, it’s 43.5× faster than DeepWalk. The quality of the generated embedding is better or competitive than the existing methods for different downstream predictions tasks.
    • Self-supervised GNNs: We develop self-supervised training algorithms and augmentation strategies for graph classification tasks.
    • Spatiotemporal Graph Learning: Our lab develops new machine learning models for the prediction of phenomena that span spatial and temporal domains and posses spatiotemporal relationships. These models are highly generalized to allow application across an array of disciplines/problems including Hydrology, Neurology, and traffic flow. We specifically focus on the construction and integration of a graphical structure that accurately captures the dependency between spatial elements across time. Graph construction/learning is integral for prediction since many spatiotemporal problem settings do not define a graphical structure or offer one that is sub-optimal for the particular prediction task.
  • Parallel Graph Clustering Algorithms
    • HipMCL: High-performance Markov Clustering 
    • Our lab designs algorithms for large-scale sparse matrix-matrix multiplications which can scale to more than 1 million hardware threads. This technique can be used in the Markov Clustering Algorithm (MCL) for large-scale network analysis.
    • Incremental Clustering
  • Machine Learning for Earth Science
    • HydroML: [Github Link][paper] We analyze historical environmental data using deep learning models to predict future extreme events including droughts and floods. We generalize our methods to enable accurate induction of hydrologic features within and across watersheds. Such induction settings allow us to make predictions of reasonable accuracy in regions that would otherwise be infeasible for analysis due to lack of historical data.
  • Graph Visualizations
    • BatchLayout: [Github Link][Paper] We develop large-scale data visualization algorithms. Most of the existing data visualization tools either consume a high amount of memory or do not utilize available cores in a computing machine i.e., they cannot fully utilize the computing resources. We aim to close this gap by introducing highly scalable algorithms.
    • BatchTree: [Github Link][Paper] BatchTree generates the edge-crossing-free and label-overlapping-free layout of a tree.
  • Scientific Computing
    • Graph Matching and Ordering