Overfitting in linear feature extraction for classification of high-dimensional image data

作者：

Highlights：

• The causes of over-fitting in feature extraction for high-dimensional datasets are revealed.

• We prove the theoretical existence of perfectly discriminative subspace projections.

• Direct, inverse relationship between the classification performance the levels of inter-class discrimination.

• Soft Discriminant Maps consistently performs better than other comparable techniques.

摘要

Highlights•The causes of over-fitting in feature extraction for high-dimensional datasets are revealed.•We prove the theoretical existence of perfectly discriminative subspace projections.•Direct, inverse relationship between the classification performance the levels of inter-class discrimination.•Soft Discriminant Maps consistently performs better than other comparable techniques.

论文关键词：Dimensionality reduction,Feature extraction,Classification,High-dimensional datasets,Overfitting

论文评审过程：Received 27 October 2014, Revised 5 November 2015, Accepted 16 November 2015, Available online 2 December 2015, Version of Record 8 February 2016.

论文官网地址：https://doi.org/10.1016/j.patcog.2015.11.015