Multiple-instance discriminant analysis

作者：

Highlights：

• We propose the MIDA algorithm for multiple-instance feature extraction.

• The MIDA algorithm can be treated as multiple-instance extension of LDA.

• MIDA can be utilized for both binary-class and multi-class learning tasks.

• MIDA can find positive prototypes and eliminate the class-label ambiguities.

• We adopt synthetic and real-world datasets to operate evaluations on MIDA.

摘要

•We propose the MIDA algorithm for multiple-instance feature extraction.•The MIDA algorithm can be treated as multiple-instance extension of LDA.•MIDA can be utilized for both binary-class and multi-class learning tasks.•MIDA can find positive prototypes and eliminate the class-label ambiguities.•We adopt synthetic and real-world datasets to operate evaluations on MIDA.

论文关键词：Multiple-instance learning,Feature extraction,Dimensionality reduction,Block coordinate ascent

论文评审过程：Received 28 June 2013, Revised 12 December 2013, Accepted 4 February 2014, Available online 13 February 2014.

论文官网地址：https://doi.org/10.1016/j.patcog.2014.02.002