SparseST: Exploiting data sparsity in spatiotemporal modeling and prediction

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Spatiotemporal data mining has a wide range of applications in modeling and predicting various complex physical systems (CPS), i.e., transportation, manufacturing, healthcare, etc. Among all the proposed methods, the Convolutional Long-Short Term Memory (ConvLSTM) has proved to be generalizable and extendable in different applications. However, ConvLSTM and its variants are computationally expensive, which makes it inapplicable in edge devices with limited computational resources. With the emerging need for edge computing in CPS, efficient AI is essential to reduce the computational cost while preserving the model performance. Common methods of efficient AI are developed to reduce model redundancy (i.e., model pruning, etc.). However, model redundancy is limited in spatiotemporal data mining as the embedded dependencies are complex and hard to capture. Instead, there is a fair level of data redundancy, which has been largely overlooked in existing research. Using the sequence of images as an example, (1) the informative pixels at each frame are usually spatially sparse (i.e., with a fixed background), and (2) the informative features in the first-order difference of two adjacent images are possibly sparse due to the slow temporal evolvement. Therefore, we develop a novel efficient ConvLSTM that pioneering exploits the sparsity of informative features. More specifically, the sparse convolution and the Delta algorithm are incorporated to exploit the spatial sparsity at each single time step and the sparsity in the first-order difference over a sequence, respectively. The experiment demonstrates that our proposed methods can achieve comparable performance to the original ConvLSTM with reduced computational complexity.