262. Exact sparsity control for multiclass linear Support Vector Machines
Invited abstract in session MC-13: Cardinality control in optimization problems for Data Science, stream Sparsity guarantee and cardinality-constrained (MI)NLPs.
Monday, 14:00-16:00Room: B100/6009
Authors (first author is the speaker)
| 1. | Pedro Duarte Silva
|
| Catolica Porto Business School, Universidade Catolica Portuguesa | |
| 2. | Immanuel Bomze
|
| Dept. of Statistics and OR, University of Vienna | |
| 3. | Laura Palagi
|
| Department of Computer, Control, and Management Engineering A. Ruberti, Sapienza University of Rome | |
| 4. | Bo Peng
|
| University of Vienna | |
| 5. | Marta Monaci
|
| Institute for System Analysis and Computer Science "Antonio Ruberti" (IASI), National Research Council of Italy | |
| 6. | Federico D'Onofrio
|
| DIAG, Sapienza University of Rome |
Abstract
Recent advances in Machine Learning have lead to the pervasiveness of AI applications that rely on highly accurate classification
algorithms. However, in many applications, because of transparency or
legal requirements, algorithm interpretability is an essential
condition that needs to be considered hand in hand with
classification accuracy. In particular case of multiclass
classification problems, while state of art deep learning neural
networks and kernel based Support Vector Machines often excel in
minimizing expected error rates, they also work as uninterpretable
black boxes that cannot be use in many problems with strict
transparency constraints. A viable alternative under these conditions
is to rely on sparse linear multiclass Support Vector Machines, that
impose a limit on the cardinality of the feature set effectively
considered. Most existing approaches to tackle this problem add
surrogate regularizers to the chosen SVM criteria, in order to reduce
the number of features used by the classifier. However, this approach
sometimes does not work as intended and, in this presentation, we
will discuss a class of sparse multiclass linear SVMs that make a
rigorous explicit control on the cardinality of feature set employed.
Numerical results on classical benchmarking datasets will be
reported, showing the efficiency and effectiveness of our approach.
Keywords
- Optimization for learning and data analysis
- Conic and semidefinite optimization
Status: accepted
Back to the list of papers