Parallel computing for Fast Spatiotemporal Weighted Regression
Parallel computing for Fast Spatiotemporal Weighted Regression
复制标题
DOI：
10.1016/j.cageo.2021.104723
复制
发表时间：
2021-03-11
影响因子：
4.4
通讯作者：
Chen, Qiyu
中科院分区：
文献类型：
--
作者：
Que, Xiang;Ma, Chao;Chen, Qiyu
时空加权回归（STWR）模型是地理加权回归（GWR）模型的扩展，用于探索时空过程的异质性。 STWR的一个关键特点是它利用之前时间阶段观察到的数据点来对最近时间阶段进行更好的拟合和预测。由于 STWR 中需要优化时间带宽和一些其他参数，因此模型校准的计算量很大。特别是当数据量很大时，STWR的校准变得非常耗时。例如，10个时间段10000个点，单核PC大约需要2307秒来处理STWR的校准。 STWR中的距离和加权矩阵都是内存密集型的，随着数据量的增加很容易导致内存不足。为了提高计算效率，我们利用消息传递接口（MPI）开发了STWR的并行计算方法。提出了 MPI 处理方法中的缓存用于校准例程。此外，还设计了矩阵分裂策略来解决内存不足的问题。我们将整体设计命名为快速 STWR (F-STWR)。在实验中，我们在高性能计算（HPC）环境中测试了F-STWR，19年来总共进行了204,611次观测。结果表明，F-STWR能够显着提高STWR处理大规模时空数据的能力。
The Spatiotemporal Weighted Regression (STWR) model is an extension of the Geographically Weighted Regression (GWR) model for exploring the heterogeneity of spatiotemporal processes. A key feature of STWR is that it utilizes the data points observed at previous time stages to make better fit and prediction at the latest time stage. Because the temporal bandwidths and a few other parameters need to be optimized in STWR, the model calibration is computationally intensive. In particular, when the data amount is large, the calibration of STWR becomes heavily time-consuming. For example, with 10,000 points in 10 time stages, it takes about 2307 s for a single-core PC to process the calibration of STWR. Both the distance and the weighted matrix in STWR are memory intensive, which may easily cause memory insufficiency as data amount increases. To improve the efficiency of computing, we developed a parallel computing method for STWR by employing the Message Passing Interface (MPI). A cache in the MPI processing approach was proposed for the calibration routine. Also, a matrix splitting strategy was designed to address the problem of memory insufficiency. We named the overall design as Fast STWR (F-STWR). In the experiment, we tested F-STWR in a High-Performance Computing (HPC) environment with a total number of 204,611 observations in 19 years. The results show that F-STWR can significantly improve STWR's capability of processing large-scale spatiotemporal data.