【論文メモ】Neural Networkを用いた特徴量の重要度の計算方法

Importance of Feature Selection for Recurrent Neural Network Based Forecasting of Building Thermal Comfort

• International Conference on Adaptive and Intelligent Systems 2014: Adaptive and Intelligent Systems pp 11-19
• https://link.springer.com/chapter/10.1007/978-3-319-11298-5_2
• https://pdfs.semanticscholar.org/adb0/6f817bf17006d01442131c31f7d2fcadecfa.pdf

概要

To select proper features tailored for particular network, we decided to use a well known sensitivity based method developed by Moody [4]. It is called Sensitivity based Pruning (SBP) algorithm. It evaluates a change in training mean squared error (MSE) that would be obtained if ith input’s influence was removed from the network. The removal of influence of input is simply modeled by replacing it by its average value.

• 特徴を一つずつ外した時に全体の誤差がどの程度変化するかを見ることで重要度を見いだす手法
  • 木系のモデルの特徴量の重要度の計算方法と似たアイディア