Using Autoencoder to Facilitate Information Retention for Data Dimension Reduction

• Main Authors: Yu-Chen Cheng, 鄭宇辰
• Other Authors: Jenq-Shiou Leu
• Format: Others
• Language: zh-TW
• Published: 2017
• Online Access: http://ndltd.ncl.edu.tw/handle/7q25x6

碩士 === 國立臺灣科技大學 === 電子工程系 === 105 === Abstract –Due to the development of internet, plentiful different data appear rapidly. The amounts of features also increase when the technology of data collecting becomes mature. Observation of different data is usually not an easy task because only a minority of people have the background knowledge of data and features. Therefore, dimensionality reduction (DR) become a familiar method to reduce the amount of features and keep the critical information. The benefits of DR are that it can sweep away useless noises, and increase the total characteristic of data. However, the loss of information during the processing of dimensionality reduction is unavoidable. When the targeted dimension is far lower than original dimension, the loss is usually too high to be endurable. To solve this problem, we use the encoder structure from autoencoder to compare with some common dimensionality reduction methods. Autoencoder is an unrestrained and flexible structure. We will use the simplest autoencoder structure as the preprocessing of Support Vector Machine (SVM) to see the result.