Learning in extreme conditions: Online and active learning with massive, imbalanced and noisy data
Başlık çevirisi mevcut değil.
- Tez No: 401678
- Danışmanlar: DR. C. LEE GILES
- Tez Türü: Doktora
- Konular: Bilgisayar Mühendisliği Bilimleri-Bilgisayar ve Kontrol, Computer Engineering and Computer Science and Control
- Anahtar Kelimeler: Belirtilmemiş.
- Yıl: 2009
- Dil: İngilizce
- Üniversite: The Pennsylvania State University
- Enstitü: Yurtdışı Enstitü
- Ana Bilim Dalı: Bilgisayar Bilimleri ve Mühendisliği Ana Bilim Dalı
- Bilim Dalı: Belirtilmemiş.
- Sayfa Sayısı: 182
Özet
Özet yok.
Özet (Çeviri)
This thesis addresses improving the performance of machine learning algorithms with a particular focus on classification tasks with large, imbalanced and noisy datasets. The field of machine learning addresses the question of how best to use experimental or historical data to discover general patterns and regularities and improve the process of decision making. However, applying machine learning algorithms to very large scale problems still faces challenges. Additionally, class imbalance and noise in the data degrade the prediction accuracy of standard machine learning algorithms. The main focus of this thesis is designing machine learning algorithms and approaches that are faster, data efficient and less demanding in computational resources to achieve scalable algorithms for large scale problems. This thesis addresses these problems in active and online learning frameworks. The particular focus of the thesis is on Support Vector Machine (SVM) algorithm with classification problems, but the proposed approaches on active and online learning are also well extensible to other widely used machine learning algorithms. This thesis first proposes a fast online Support Vector Machine algorithm (LASVM) that has an outstanding speed improvement over the classical (batch) SVM and other online SVM algorithms, while preserving the classification accuracy rates of the state-of-the-art SVM solvers. The ability to handle streaming data, speed improvement in both training and recognition and the demand for less memory with the online learning setting enable SVM to be applicable to very large datasets. The effectiveness of LASVM and active learning in real world problems is assessed by targeting the name disambiguation problem in CiteSeer's repository of scientific publications. The thesis then presents an efficient active learning framework to target the expensive labeling problem for producing training data. The proposed method yields an efficient querying system and removes the barriers of applying active learning to very large scale datasets due to the high computational costs. We then point out that even when the labels are readily available, active learning can still be used to reach out to the most informative instances in the training data. By applying active sample selection and early stopping in the online SVM, we show that the algorithm can reach and even exceed the prediction accuracies of baseline setting of LASVM with random sampling. Our experimental results also reveal that active learning can be a highly effective method for dealing with the class imbalance problem. We further propose a hybrid method of oversampling and active learning to form an adaptive technique (named VIRTUAL) to efficiently resample the minority class instances in imbalanced data classification. Finally, we propose a non-convex online SVM algorithm (LASVM-NC) based on the Ramp loss, which has strong ability of suppressing the in uences of outliers in noisy datasets. Then, again in the online learning setting, we propose an outlier filtering mechanism that approximates non-convex behavior in convex optimization (LASVM-I). These two algorithms are built on an online SVM solver (LASVM-G) which leverages the duality gap to obtain more trustworthy intermediate models. Our experimental results show that the proposed approaches yield a more scalable online SVM algorithm with sparser models and less computational running time both in the training and recognition phases without sacrificing generalization performance. In the end, we also point out the relation between the non-convex behavior in SVMs and active learning.
Benzer Tezler
- Potentialities for and limits to inclusion by education: The case of Syrian children's education in Turkey and child labour
Eğitim tarafından içermede potansiyeller ve limitler: Türkiye'deki Suriyeli çocukların eğitimi ve çocuk işçiliği
YASEMİN KIZILOĞLU
Yüksek Lisans
İngilizce
2021
Sosyal HizmetOrta Doğu Teknik ÜniversitesiSosyal Politika Ana Bilim Dalı
DR. ÖĞR. ÜYESİ MEHMET OKYAYUZ
- Sequential nonlinear learning
Ardışık doğrusal olmayan öğrenme
NURİ DENİZCAN VANLI
Yüksek Lisans
İngilizce
2015
Elektrik ve Elektronik Mühendisliğiİhsan Doğramacı Bilkent ÜniversitesiElektrik-Elektronik Mühendisliği Ana Bilim Dalı
DOÇ. SÜLEYMAN SERDAR KOZAT
- Güç kalitesi olaylarının makine öğrenme teknikleri ile sınıflandırılması
Classification of power quality events using machine learning methods
FERHAT UÇAR
Doktora
Türkçe
2018
Elektrik ve Elektronik MühendisliğiFırat ÜniversitesiElektrik-Elektronik Mühendisliği Ana Bilim Dalı
DR. ÖĞR. ÜYESİ FİKRET ATA
PROF. DR. BEŞİR DANDIL
- Geliştirilmiş rastgele vektör işlevsel bağlantı ağları ile dağıtım şebekelerinde arıza türü ve yerinin tespiti
Fault type and location detection in distribution networks with improved random vector functional link networks
CEM HAYDAROĞLU
Doktora
Türkçe
2022
Elektrik ve Elektronik MühendisliğiDicle ÜniversitesiElektrik ve Elektronik Mühendisliği Ana Bilim Dalı
DOÇ. DR. BİLAL GÜMÜŞ
- Design and performance evaluation of demand forecasting system for online food data
Sanal yemek verisi üzerinde talep tahmin sistemi tasarımı ve başarım değerlendirmesi
MELTEM ARSLAN
Yüksek Lisans
İngilizce
2024
Bilgisayar Mühendisliği Bilimleri-Bilgisayar ve KontrolGalatasaray ÜniversitesiBilgisayar Mühendisliği Ana Bilim Dalı
DOÇ. DR. GÜLFEM ALPTEKİN