Kolmogorov-Arnold Network Based Feature Extraction of Support Vector Machines
Acta Polytechnica Hungarica, cilt.23, sa.7, ss.161-183, 2026 (SCI-Expanded, Scopus)
- Yayın Türü: Makale / Tam Makale
- Cilt numarası: 23 Sayı: 7
- Basım Tarihi: 2026
- Doi Numarası: 10.12700/aph.23.7.2026.7.9
- Dergi Adı: Acta Polytechnica Hungarica
- Derginin Tarandığı İndeksler: Science Citation Index Expanded (SCI-EXPANDED), Scopus
- Sayfa Sayıları: ss.161-183
- Anahtar Kelimeler: AISA, Feature Extraction, Kolmogorov-Arnold Representation, Support Vector Machine
- Yıldız Teknik Üniversitesi Adresli: Evet
Özet
Support Vector Machines (SVMs) have been applied to the small datasets using pre-processing methods since their learning-based feature extraction capabilities are limited compared to deep neural networks. In this study, a Kolmogorov-Arnold Network (KAN) is used for feature extraction to achieve a high approximation performance of SVMs. The proposed pre-processing layer involves optimization of functions and scaling parameters using the Adolescent Identity Search Algorithm (AISA). First, trigonometric basis functions are optimized with constant parameters. Then, the parameters of these functions are optimized to construct an optimal Kolmogorov-Arnold network using a smaller number of basis functions. In the proposed model, the internal layer functions of the KAN become trigonometric functions, while the external layer functions are Gaussian distribution functions in the kernel-induced feature space of the SVM model. At the same time, the hyperparameters of the SVM model, V C dimension, ε, and σ of the kernel functions are optimized for further performance improvement. The signal properties of the input and feature layers such as energy, central frequency, entropy, correlation, principal components, and clusters are analyzed to discuss the information content of the layers. The variation in feature signal properties provides important evidence in the KAN regression layer literature regarding how the classification and regression performance is improved using the KAN layer. The proposed KAN-SVM model has been applied for the regression and classification of benchmark datasets. The resulting KAN layer and hyperparameters for the identification of a nonlinear dynamic system are discussed by considering the mathematical model of the system. In addition, input-output data recorded from a newly_developed real-time experimental setup has been accurately identified using the proposed KAN-SVM model.