Scalable geographically weighted regression for big data. GeoComputation 2019
conference contributionposted on 01.12.2019 by Daisuke Murakami, Narumasa Tsutsumida, Takahiro Yoshida, Tomoki Nakaya, Binbin Lu
Any type of content contributed to an academic conference, such as papers, presentations, lectures or proceedings.
This study develops a scalable GWR (ScaGWR) for large samples. Unlike existing GWR algorithms, the ScaGWR achieves a linear time estimation through pre-compression of large matrices and vectors before the leave-one-out cross validation (LOOCV), which is the heaviest part in the standard GWR. Our proposed algorithm enables us to estimate a regularized GWR from one million samples even without parallelization. The R code is available in the R package scgwr. Besides, for faster computation, the code is embedded into C++ code and implemented in the GWmodel package.