Exhaustive Search of Feature Subsets for Support Vector Machine Classification

• Jun Kitazono

• Kenji Nagata

• Shinichi Nakajima

• Akira Manda

• Satoshi Eifuku

• Ryoi Tamura

• Masato Okada

抄録

Feature selection in machine learning is an important process for improving the generalization capability and interpretability of learned models through the selection of a relevant feature subset. In the last two decades, a number of feature selection methods, such as L1 regularization and automatic relevance determination have been intensively developed and used in a wide range of areas. We can select a relevant subset of features, by using these feature selection methods. In this study, we apply an exhaustive search, instead of these methods, to the neural data recorded in the area of brain involved in face recognition. We evaluate how accurately every subset of recorded neurons can discriminate faces, by using SVM classifiers and cross validation. We show that there are a number of highly accurate neuron subsets. All of these results demonstrate that we should not select only one feature subset but exhaustively evaluate every feature subset.Feature selection in machine learning is an important process for improving the generalization capability and interpretability of learned models through the selection of a relevant feature subset. In the last two decades, a number of feature selection methods, such as L1 regularization and automatic relevance determination have been intensively developed and used in a wide range of areas. We can select a relevant subset of features, by using these feature selection methods. In this study, we apply an exhaustive search, instead of these methods, to the neural data recorded in the area of brain involved in face recognition. We evaluate how accurately every subset of recorded neurons can discriminate faces, by using SVM classifiers and cross validation. We show that there are a number of highly accurate neuron subsets. All of these results demonstrate that we should not select only one feature subset but exhaustively evaluate every feature subset.

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） 2013 (8), 1-6, 2013-02-20

詳細情報

• CRID: 1571980077826192640
• NII論文ID: 110009550156
• NII書誌ID: AN10505667
• 本文言語コード: en
• データソース種別:
  • CiNii Articles