機器學習之特徵選擇-基於正交實驗的探討

Abstract

在機器學習領域當中，特徵選擇一直以來都是一個重要的課題，尤其以行為辨識(Activity Recognition)而言，我們利用許多不同的感應器來擷取各種大量的資訊，假如能透過特徵選擇的技術來挑選出重要特徵，將有許多好處，例如增快辨識速度、提高辨識準確度等等。本論文提出一個基於正交實驗的特徵選擇法，並以循序收納選擇法(Sequential Forward Selection)為基準，比較並探討此法的優劣與適用性。

Feature selection is an important issue in the problem of machine learning. Especially in the domain of activity recognition, many researchers try to make use of multiple heterogeneous sensors and thus receive a large amount of signals. Many features can be extracted, hence feature selection becomes more important. In this thesis, we propose a feature selection method based on orthogonal experimental design and compare this method with equential forward feature selection in terms of the accuracy of the model induced by the selected feature subset versus the number of treatments and the number of selected features.