Rapid Feature Selection Based on Random Forests for High-Dimensional Data
この論文をさがす
抄録
One of the important issues of machine learning is obtaining essential information from high-dimensional data for discrimination. Dimensionality reduction is a means to reduce the burden of dimensionality due to large-scale data. Feature selection determines significant variables and contributes to dimensionality reduction. In recent years, the random forests method has been the focus of research because it can perform appropriate variable selection even with high-dimensional data holding high correlations between dimensionality. There exist many feature selection methods based on random forests. These methods can appropriately extract the minimum subset of important variables. However, these methods need more computation time than the original random forests method. An advantage of the random forests method is its speed. Therefore, this paper aims to propose a rapid feature selection method for high-dimensional data. Rather than searching the minimum subset of important variables, our method aims to select meaningful variables quickly under the assumption that the number of variables to be selected is determined beforehand. Two main points are introduced to enable faster calculations. One is reduction in the calculation time of weak learners. The other is adopting two types of feature selection: "filter" and "wrapper." In addition, although most present methods use only "mean decrease accuracy," we calculate the magnitude of features by combining "mean decrease accuracy" and "Gini importance." As a result, our method can reduce computation time in cases where generated trees have many nodes. More specifically, our method can reduce the number of important variables to 0.8% on an average without losing the information for classification. In conclusion, our proposed method based on random forests is found to be effective for achieving rapid feature selection.One of the important issues of machine learning is obtaining essential information from high-dimensional data for discrimination. Dimensionality reduction is a means to reduce the burden of dimensionality due to large-scale data. Feature selection determines significant variables and contributes to dimensionality reduction. In recent years, the random forests method has been the focus of research because it can perform appropriate variable selection even with high-dimensional data holding high correlations between dimensionality. There exist many feature selection methods based on random forests. These methods can appropriately extract the minimum subset of important variables. However, these methods need more computation time than the original random forests method. An advantage of the random forests method is its speed. Therefore, this paper aims to propose a rapid feature selection method for high-dimensional data. Rather than searching the minimum subset of important variables, our method aims to select meaningful variables quickly under the assumption that the number of variables to be selected is determined beforehand. Two main points are introduced to enable faster calculations. One is reduction in the calculation time of weak learners. The other is adopting two types of feature selection: "filter" and "wrapper." In addition, although most present methods use only "mean decrease accuracy," we calculate the magnitude of features by combining "mean decrease accuracy" and "Gini importance." As a result, our method can reduce computation time in cases where generated trees have many nodes. More specifically, our method can reduce the number of important variables to 0.8% on an average without losing the information for classification. In conclusion, our proposed method based on random forests is found to be effective for achieving rapid feature selection.
収録刊行物
-
- 研究報告数理モデル化と問題解決(MPS)
-
研究報告数理モデル化と問題解決(MPS) 2012 (3), 1-7, 2012-07-09
- Tweet
詳細情報 詳細情報について
-
- CRID
- 1574231877590930944
-
- NII論文ID
- 110009421219
-
- NII書誌ID
- AN10505667
-
- 本文言語コード
- en
-
- データソース種別
-
- CiNii Articles