Publications

Randomized Algorithms for Symmetric Nonnegative Matrix Factorization

Published in SIAM Journal on Matrix Analysis and Applications, 2024

We present the first two algorithms for randomized Symmetric Nonnegative Matrix Factorization. Our methods lead to 5.5-7.5x speed up on large graph datasets and preserve found clustering quality. Additionally, we prove theoretical results for our methods, guaranteeing accuracy with high probability.

Recommended citation: Hayashi, Koby, Aksoy, Sinan G., Ballard, Grey, et al., "Randomized Algorithms for Symmetric Nonnegative Matrix Factorization," SIAM Journal on Matrix Analysis and Applications 46, no. 1 (2025), https://doi.org/10.1137/24m1638355. http://hayakb95.github.io/files/BigData_2022.pdf

Skew Symmetric Adjacency Matrices for Clustering Directed Graphs

Published in 2022 IEEE International Conference on Big Data (Big Data), 2022

We develope a spectral algorithm for clustering directed Graphs. The method uses spectal properties of the Skew-Symmetric Adjacency matrix to find clusters with large imbalanced cuts.

Recommended citation: K. Hayashi, S. G. Aksoy and H. Park, "Skew-Symmetric Adjacency Matrices for Clustering Directed Graphs," 2022 IEEE International Conference on Big Data (Big Data), Osaka, Japan, 2022, pp. 555-564, doi: 10.1109/BigData55660.2022.10020413. http://hayakb95.github.io/files/BigData_2022.pdf

PLANC: Parallel Low Rank Approximation with Non-negativity Constraints

Published in ACM Transactions on Mathematical Software, 2021

PLANC is a software package for computing low-rank tensor and matrix approximations with nonnegativity constraints. The package supports shared and distributed memory algorithms. Sparse and dense NMF are supported along with Dense CP of general tensors.

Recommended citation: Srinivas Eswar, Koby Hayashi, Grey Ballard, Ramakrishnan Kannan, Michael A. Matheson, and Haesun Park. 2021. PLANC: Parallel Low-rank Approximation with Nonnegativity Constraints. ACM Trans. Math. Softw. 47, 3, Article 20 (September 2021), 37 pages. https://doi.org/10.1145/3432185 http://hayakb95.github.io/files/TOMS_PLANC.pdf

Distributed-Memory Parallel Symmetric Nonnegative Matrix Factorization

Published in SC20: International Conference for High Performance Computing, Networking, Storage and Analysis, 2020

We present two parallel distributed memory algorithms for finding approximate solutions to the Symmetric Nonnegative Matrix Factorization problem. These are the first such algorithms. One is based of the Alternating Nonnegative Least Squares method. The other is based on the Gauss-Newton method with Conjugate Gradients.

Recommended citation: S. Eswar, K. Hayashi, G. Ballard, R. Kannan, R. Vuduc and H. Park, "Distributed-Memory Parallel Symmetric Nonnegative Matrix Factorization," SC20: International Conference for High Performance Computing, Networking, Storage and Analysis, Atlanta, GA, USA, 2020, pp. 1-14, doi: 10.1109/SC41405.2020.00078. http://hayakb95.github.io/files/SymNMF_SC20.pdf

Hypergraph Random Walks, Laplacians, and Clustering

Published in CIKM20: Proceedings of the 29th ACM International Conference on Information and Knowledge Management, 2020

We propose a flexible framework for clustering hypergraph-structured data based on recently proposed random walks utilizing edge-dependent vertex weights.

Recommended citation: Koby Hayashi, Sinan G. Aksoy, Cheong Hee Park, and Haesun Park. 2020. Hypergraph Random Walks, Laplacians, and Clustering. In Proceedings of the 29th ACM International Conference on Information & Knowledge Management (CIKM20). Association for Computing Machinery, New York, NY, USA, 495–504. https://doi.org/10.1145/3340531.3412034 http://hayakb95.github.io/files/Hypergraph_RW_CIKM20.pdf

Parallel Nonnegative CP Decomposition of Dense Tensors

Published in 2018 IEEE 25th International Conference on High Performance Computing (HiPC), 2018

We present a distributed memory parallel algorithm for computing an approximate nonnegative CP decomposition of dense tensors. The algorithm uses dimensions trees to reduce computation and a tuned communication grid to reduce communication.

Recommended citation: G. Ballard, K. Hayashi and K. Ramakrishnan, "Parallel Nonnegative CP Decomposition of Dense Tensors," 2018 IEEE 25th International Conference on High Performance Computing (HiPC), Bengaluru, India, 2018, pp. 22-31, doi: 10.1109/HiPC.2018.00012. http://hayakb95.github.io/files/NTF_HiPC18.pdf

Shared-Memory Parallelization of MTTKRP for Dense Tensors

Published in PPoPP "18: Proceedings of the 23rd ACM SIGPLAN Symposium on Principles and Practice of Parallel Programming (Short Paper), 2018

Matricized Tensor Times Khatri-Rao Product (MTTKRP) is the main computational kernel when computing a CP Decomposition of tensors. This paper describes parallelization strategies for efficiently computing MTTKRP for dense tensors.

Recommended citation: Hayashi, Koby. (2018). "Shared-Memory Parallelization of MTTKRP for Dense Tensors." Proceedings of the 23rd ACM SIGPLAN Symposium on Principles and Practice of Parallel Programming. http://hayakb95.github.io/files/Sharmed_Mem_MTTKRP_Full.pdf

Dynamic functional connectivity and individual differences in emotions during social stress

Published in Human Brain Mapping, 2017

Exposure to acute stress induces multiple emotional responses, each with their own unique temporal dynamics. Dynamic functional connectivity (dFC) measures the temporal variability of network synchrony and captures individual differences in network neurodynamics. This study investigated the relationship between dFC and individual differences in emotions induced by an acute psychosocial stressor.

Recommended citation: Tobia, M. J., Hayashi, K., Ballard, G., Gotlib, I. H., & Waugh, C. E. (2017). Dynamic functional connectivity and individual differences in emotions during social stress. Human Brain Mapping, 38(12), 6185–6205. https://doi.org/10.1002/hbm.23821 http://hayakb95.github.io/files/Human_Brain_Mapping.pdf