1130. Strict control of sparsity for classification problems
Invited abstract in session WC-35: Cardinality-constrained optimization with guarantees, stream Continuous and mixed-integer nonlinear programming: theory and algorithms.
Wednesday, 12:30-14:00Room: Michael Sadler LG15
Authors (first author is the speaker)
| 1. | Immanuel Bomze
|
| Dept. of Statistics and OR, University of Vienna | |
| 2. | Federico D'Onofrio
|
| DIAG, Sapienza University of Rome | |
| 3. | Pedro Duarte Silva
|
| Catolica Porto Business School, Univ. Catolica Portuguesa | |
| 4. | Marta Monaci
|
| Universitas Mercatorum | |
| 5. | Laura Palagi
|
| Department of Computer, Control, and Management Engineering A. Ruberti, Sapienza University of Rome | |
| 6. | Bo Peng
|
| University of Southern California |
Abstract
To ensure explainability in AI and transparency in Machine Learning, control of sparsity is crucial. However, most of the popular approaches use surrogate regularizers to reduce the number of variables involved in the classifier, which sometimes does not work as intended. Instead, we follow a recent strain of research by rigorous control of this number directly, employing an explicit cardinality constraint on the original features. The resulting nonconvex QCQP is NP-hard and we propose conic relaxations to tackle it.
Keywords
- Continuous Optimization
- Global Optimization
- Programming, Mixed-Integer
Status: accepted
Back to the list of papers