Separability-Oriented Subclass Discriminant Analysis
- Submitting institution
-
The University of West London
- Unit of assessment
- 11 - Computer Science and Informatics
- Output identifier
- 11041
- Type
- D - Journal article
- DOI
-
10.1109/TPAMI.2017.2672557
- Title of journal
- IEEE Transactions on Pattern Analysis and Machine Intelligence
- Article number
- -
- First page
- 409
- Volume
- 40
- Issue
- 2
- ISSN
- 0162-8828
- Open access status
- Technical exception
- Month of publication
- -
- Year of publication
- 2017
- URL
-
-
- Supplementary information
-
-
- Request cross-referral to
- -
- Output has been delayed by COVID-19
- No
- COVID-19 affected output statement
- -
- Forensic science
- No
- Criminology
- No
- Interdisciplinary
- No
- Number of additional authors
-
3
- Research group(s)
-
-
- Citation count
- 16
- Proposed double-weighted
- No
- Reserve for an output with double weighting
- No
- Additional information
- Linear discriminant analysis (LDA) is a classical method for discriminative dimensionality reduction. LDA projects data in such a way that maximises within-class compactness and between-class separation. LDA has been extended to consider subclasses with classes, but existing extensions produce subclasses that are significantly overlapping. This paper presents a new extension, SSDA, that additionally maximises between-subclass separation. Experiments show that SSDA has better classification performance than LDA and its subclass-based extensions, and that classification performance is highly correlated with class separation. This finding has inspired the design of new loss functions for deep learning, and some subsequent papers have been published.
- Author contribution statement
- -
- Non-English
- No
- English abstract
- -