Random projections as regularizers: learning a linear discriminant from fewer observations than dimensions
- Submitting institution
-
The University of Birmingham
- Unit of assessment
- 11 - Computer Science and Informatics
- Output identifier
- 24112693
- Type
- D - Journal article
- DOI
-
10.1007/s10994-014-5466-8
- Title of journal
- Machine Learning
- Article number
- -
- First page
- 257
- Volume
- 99
- Issue
- 2
- ISSN
- 0885-6125
- Open access status
- Out of scope for open access requirements
- Month of publication
- August
- Year of publication
- 2014
- 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
-
1
- Research group(s)
-
-
- Citation count
- 15
- Proposed double-weighted
- No
- Reserve for an output with double weighting
- No
- Additional information
- We prove that an averaging ensemble of randomly-projected FLD classifiers succeeds in high-dimensional low-sample settings, where an existing alternative provably fails. We establish a theoretical link between the strong-classifier and weak-classifier-ensemble pair for the first time, yielding an efficient algorithm and significant
accuracy gains. An early version won the best paper award of ACML'13.
The student author obtained academic lectureship. An invited paper (Kaban, Analysis & Applications 2020) proves the ensemble size needs only be linear in data dimension to recover all parameters. It impacted
statistical methodology (https://rss.onlinelibrary.wiley.com/doi/full/10.1111/rssb.12228),
signal processing (https://ieeexplore.ieee.org/abstract/document/7542135), medical
research (https://link.springer.com/chapter/10.1007/978-3-319-47106-8_7),
and evolutionary algorithms (https://link.springer.com/chapter/10.1007/978-3-319-47106-8_7).
- Author contribution statement
- -
- Non-English
- No
- English abstract
- -