Seminario: “The Twigs Classifier: A Novel Supervised, Non-Parametric, and Adaptive Approach for High-Accuracy Classification”

Segnaliamo il seminario dal titolo The Twigs Classifier: A Novel Supervised, Non-Parametric, and Adaptive Approach for High-Accuracy Classification, che si svolgerà, nell'ambito dell'insegnamento in Neural and Rehabilitation Engineering, lunedì 9 novembre alle ore 12:00 in aula N22.

The Twigs Classifier: A Novel Supervised, Non-Parametric, and Adaptive Approach for High-Accuracy Classification
Kamel Mebarkia 1, *, Aicha Reffad 2
1 LIS Laboratory, Electronics Department, Faculty of Technology, Setif 1 University-Ferhat ABBAS, 19000 Setif, Algeria.
E-mail: Link identifier #identifier__67932-1kamel.mebarkia@rwth-aachen.de
2 LAS Laboratory, Electrotechnics Department, Faculty of Technology, Setif 1 University-Ferhat ABBAS, 19000 Setif, Algeria. E-mail: Link identifier #identifier__145562-2reffad.aicha@gmail.com
Abstract
In this work, we introduce the twigs classifier, a new supervised, non-parametric, and adaptive classification method. The core idea behind this classifier is the concept of twigs, which are defined as boundary vectors (BVs) along with their corresponding twin support vectors (SVs). These twigs are identified through a novel, intuitive, and efficient algorithm called the boundary vector bisection-based algorithm (BVB).
The BVB algorithm iteratively drives a population of scattered seeds toward the decision boundaries between classes, operating independently of both the number of classes and the dimensionality of the data. We also identify and address several limitations of the original BVB, leading to an enhanced version the BVB circle-based algorithm (BVBC), specifically designed to handle challenging scenarios such as spiral-shaped class distributions.
Classification with the twigs classifier is remarkably simple: it computes a dot product between the nearest twig to the target object and the twig-object vector. Through adaptive orientation of twigs and strategic pruning, we achieve significant gains in classification accuracy (CA).
The BVB/BVBC algorithms and the twigs classifier are rigorously evaluated using both synthetic datasets and 20 real-world datasets from the UCI repository. Our results demonstrate the classifier’s robustness in multi-class, imbalanced, and high-dimensional settings. Notably, the twigs classifier outperforms most state-of-the-art classifiers  including deep learning in several cases and achieves a misclassification rate below 1% across the majority of tested scenarios.

Keywords: Machine learning, Supervised learning, Boundary vectors, Support vectors, Bisection principle, Data reduction.