[S-TT40] 最先端ベイズ統計学が拓く地震ビッグデータ解析

[STT40-P01] Fast and high-precision estimation for temporal variation of aftershocks immediately after a main shock with Gaussian process regression

*森川 耕輔¹、長尾 大道^2,3、伊藤 伸一^2,3、酒井 慎一^2,5、平田 直^2,4 (1.大阪大学、2.東京大学地震研究所、3.東京大学大学院情報理工学系研究科、4.防災科学技術研究所、5.東京大学大学院情報学環・学際情報学府)

キーワード：統計地震学、ガウス過程回帰、Noise contrastive estimation

A large main shock often makes it challenging to identify a number of subsequent aftershocks, and it distorts estimates for the distribution of magnitudes and arrival times of the aftershocks. We early proposed a method to correct the bias with high accuracy by incorporating a detection function with the Gaussian process regression, a flexible nonparametric Bayesian method. However, the algorithm estimating hyperparameters takes much time due to the MCMC (Markov chain Monte Carlo) sampling and needs to know tuning parameters beforehand. In this talk, we propose a fast estimation algorithm to estimate the hyperparameters by using the technique of noise contrastive estimation ( NCE). We apply our proposed algorithm to various earthquake catalogs and show that the proposed method can estimate hyperparameters fast in keeping the high precision for estimating the target parameters.