310. Anomaly Detection Using the Cloud of Spheres Classification Method
Invited abstract in session WB-4: Optimization and learning for estimation problems, stream Optimization for machine learning.
Wednesday, 10:30-12:30Room: B100/5013
Authors (first author is the speaker)
| 1. | Paula Amaral
|
| Faculdade de Ciências e Tecnologia, Universidade Nova de Lisboa | |
| 2. | Tiago Dias
|
| Department of Mathematics, NOVA School of Science and Technology | FCT NOVA |
Abstract
Machine learning models have been extensively applied across various domains, often without a deep understanding of their underlying mechanisms. Black-box models, such as Deep Neural Networks, present significant challenges in counterfactual analysis, interpretability, and explainability.
In this presentation, we introduce a novel binary classification model called the Connected Cloud of Spheres. This model is formulated as a Mixed-Integer Nonlinear Programming (MINLP) problem, aiming to minimize the number of spheres required to accurately classify data points. The method is particularly effective for datasets with highly non-linear and non-convex structures while remaining adaptable to linearly separable cases. Unlike neural networks, our approach operates directly in the original feature space, eliminating the need for kernel functions or extensive hyperparameter tuning.
Although primarily designed for binary classification, this method can be extended for anomaly detection, particularly in scenarios where negative examples are unavailable at the outset. Additionally, we discuss heuristic strategies for outlier identification and explainability, offering insights into how this approach enhances model transparency and interpretability.
Keywords
- Optimization for learning and data analysis
- Nonlinear mixed integer optimization
Status: accepted
Back to the list of papers