The team of Tohoku University attended sub-competition 2b in the competition on spatial statistics for large datasets, where prediction on 100,000 testing points were to be constructed conditional on 900,000 training points. We chose a covariance tapering approach in a simplified way to manage one million spatial data points. Dividing \([0,1]^2\) into \(30\times 30\) sub-regions with equal area, we construct predictors separately in each sub-region conditional on training data over the extended sub-region with length enlarged by \(\sqrt{2}\) by fitting Matérn class covariances.

中文翻译：

---

大数据空间统计竞争探讨

东北大学的团队参加了大型数据集空间统计竞赛的分赛 2b，其中以 900,000 个训练点为条件构建了对 100,000 个测试点的预测。我们以一种简化的方式选择了协方差逐渐变细的方法来管理一百万个空间数据点。将\([0,1]^2\)划分为\(30\times 30\)个等面积的子区域，我们在每个子区域中分别构建预测器，条件是在长度扩大的扩展子区域上的训练数据通过\(\sqrt{2}\)拟合 Matérn 类协方差。