以核密度為基礎推估分類器之預測機

Abstract

本論文比較幾種核心密度估計方法(Kernel Density Estimation)並提出可變帶寬核密 度估計(Adaptive Density Estimation)的挑選模式。核密度估計是一種無母數統計方 法，相對有母數統計較不受特定框架影響，有較高的彈性和配適性，而可變帶寬 核密度估計較固定帶寬核密度估計有更佳的配適性。論文中討論此兩種核密度估 計方法，並提出 RVKDE 的帶寬優化演算法 Elevated RVKDE，以減少需調整的參 數，在多種人工合成資料集上實驗，結果顯示此方法在大部分情況下表現優於其 他方法;最後文中介紹如何應用密度估計於分類器的機率估計，及應用於實際登 革熱資料集，並和其他分類演算法比較分類能力。

This study compares kernel density estimation (KDE) algorithms which is a branch of nonparametric statistics and propose a method to optimize bandwidth selection in Relaxed Variable Kernel Density Estimation (RVKDE) called Elevated RVKDE. KDE methods have fixed KDE and adaptive KDE. Nonparametric method is flexible and adaptive KDEs have even better goodness of fit than fixed KDEs. However, RVKDE is an adaptive method that have to tune a smoothing parameter to reach the ideal condition. In this study, we propose the method that there’s no need to tune this smoothing parameter anymore and outperforms other KDE methods mostly on synthesis dataset. This study also introduce how to implement density estimation to a probability estimated classifier and compare the performance of it with several machine learning algorithms on the dengue dataset.