Dengesiz Veri Setleri Üzerinde Sınıflandırma İçin Alternatif Bir Özellik Seçimi Yaklaşımı

Tezin Türü: Yüksek Lisans

Tezin Yürütüldüğü Kurum: Dokuz Eylül Üniversitesi, Fen Bilimleri Enstitüsü, Veri Bilimi, Türkiye

Tezin Onay Tarihi: 2026

Tezin Dili: Türkçe

Öğrenci: İhsan Mutlu Karadeniz

Danışman: Özlem Ege Oruç

Özet:

In imbalanced datasets, standard classification algorithms are biased toward the

majority class, which leads to insufficient representation of minority classes and

significant performance degradation in critical applications such as medical diagnosis,

cybersecurity, and financial fraud detection. DDSC-SMOTE, which addresses this

issue, provides an effective oversampling approach by generating distribution-guided

synthetic samples; however, its dense spectral clustering component introduces

substantial computational and memory overhead, resulting in poor scalability for

large-scale datasets.

This thesis preserves the core structure of DDSC-SMOTE while redesigning

its clustering component to improve scalability. Instead of conventional spectral

clustering, which relies on dense similarity matrices, scalable U-SPEC- and

U-SENC-based clustering methods are integrated. Accordingly, two novel variants,

DDUSPEC-SMOTE and DDUSENC-SMOTE, are proposed. The proposed methods

are evaluated against DDSC-SMOTE and widely used oversampling techniques on

datasets with varying scales and imbalance ratios.

Experimental results demonstrate that the proposed methods achieve performance

comparable to DDSC-SMOTE on small and medium-scale datasets while preserving

the underlying learning behavior. On the large-scale Covertype dataset, where

DDSC-SMOTE cannot be executed due to memory limitations, the proposed

approaches operate successfully, significantly improving scalability. Furthermore,

component-level analysis shows that after improving scalability, the main

computational cost shifts toward the filtering stage, and certain mechanisms do

not consistently contribute across all datasets.Overall, this study proposes two scalable variants that maintain the effectiveness of

DDSC-SMOTE while substantially improving its applicability to large-scale datasets.

Keywords: Imbalanced data, oversampling, scalable spectral clustering,

DDSC-SMOTE, U-SPEC, U-SENC