AGCRS: An Adaptive Generalized Storage Scheme for Large Sparse Tensors

Sparse tensor storage is critical for efficient processing of high-dimensional data in scientific computing and machine learning applications. Existing formats for storing large tensors often face trade-offs between compression efficiency, scalability, and construction cost, especially for tensors with varying dimensionality and sparsity patterns. This paper proposes a novel tensor storage format, namely, Adaptive Generalized Compressed Row Storage (AGCRS), that employs a cost-guided partitioning strategy to adaptively compress sparse tensors without manual tuning. AGCRS achieves superior compression ratios compared to state-of-the-art formats such as ALTO, CSF, and HiCOO, demonstrating consistent performance across real-world and synthetic tensors up to 16 dimensions. Our evaluation shows that AGCRS balances compression gains with reasonable construction time and scales efficiently with tensor order, making it a promising storage format for large-scale, high-order sparse tensor applications.