Integrated Fisher linear discriminants: An empirical study
作者：
Highlights：
• An optimal threshold is found out from a series of empirical threshold formulas developed for Fisher linear discriminants based on classification accuracies.
• Weight vectors and thresholds are updated by an epoch-limited iterative learning strategy.
• The singular within-class scatter matrices are reduced in dimensionality but not added with perturbations.
• A coding system enlarges class margins and approximately preserves neighborhood relationships.
• An integrated learning algorithm improves the learning and generalization performances of Fisher linear discriminants.
摘要
•An optimal threshold is found out from a series of empirical threshold formulas developed for Fisher linear discriminants based on classification accuracies.•Weight vectors and thresholds are updated by an epoch-limited iterative learning strategy.•The singular within-class scatter matrices are reduced in dimensionality but not added with perturbations.•A coding system enlarges class margins and approximately preserves neighborhood relationships.•An integrated learning algorithm improves the learning and generalization performances of Fisher linear discriminants.
论文关键词：Fisher linear discriminants,Imbalanced datasets,Empirical thresholds,Neighborhood-preserving transformations,Iterative learning
论文评审过程：Received 6 December 2012, Revised 16 July 2013, Accepted 26 July 2013, Available online 9 August 2013.
论文官网地址：https://doi.org/10.1016/j.patcog.2013.07.021
