Date of Award
August 2020
Document Type
Dissertation
Degree Name
Doctor of Philosophy (PhD)
Department
School of Computing
Committee Member
Ilya Safro
Committee Member
Amy Apon
Committee Member
Alexander Herzog
Committee Member
Brian Dean
Committee Member
Lewis Frey
Abstract
The continuous increase in the size of datasets introduces computational challenges for machine learning algorithms. In this dissertation, we cover the machine learning algorithms and applications in large-scale data analysis in manufacturing and healthcare. We begin with introducing a multilevel framework to scale the support vector machine (SVM), a popular supervised learning algorithm with a few tunable hyperparameters and highly accurate prediction. The computational complexity of nonlinear SVM is prohibitive on large-scale datasets compared to the linear SVM, which is more scalable for massive datasets. The nonlinear SVM has shown to produce significantly higher classification quality on complex and highly imbalanced datasets. However, a higher classification quality requires a computationally expensive quadratic programming solver and extra kernel parameters for model selection. We introduce a generalized fast multilevel framework for regular, weighted, and instance weighted SVM that achieves similar or better classification quality compared to the state-of-the-art SVM libraries such as LIBSVM. Our framework improves the runtime more than two orders of magnitude for some of the well-known benchmark datasets. We cover multiple versions of our proposed framework and its implementation in detail. The framework is implemented using PETSc library which allows easy integration with scientific computing tasks. Next, we propose an adaptive multilevel learning framework for SVM to reduce the variance between prediction qualities across the levels, improve the overall prediction accuracy, and boost the runtime. We implement multi-threaded support to speed up the parameter fitting runtime that results in more than an order of magnitude speed-up. We design an early stopping criteria to reduce the extra computational cost when we achieve expected prediction quality. This approach provides significant speed-up, especially for massive datasets. Finally, we propose an efficient low dimensional feature extraction over massive knowledge networks. Knowledge networks are becoming more popular in the biomedical domain for knowledge representation. Each layer in knowledge networks can store the information from one or multiple sources of data. The relationships between concepts or between layers represent valuable information. The proposed feature engineering approach provides an efficient and highly accurate prediction of the relationship between biomedical concepts on massive datasets. Our proposed approach utilizes semantics and probabilities to reduce the potential search space for the exploration and learning of machine learning algorithms. The calculation of probabilities is highly scalable with the size of the knowledge network. The number of features is fixed and equivalent to the number of relationships or classes in the data. A comprehensive comparison of well-known classifiers such as random forest, SVM, and deep learning over various features extracted from the same dataset, provides an overview for performance and computational trade-offs. Our source code, documentation and parameters will be available at https://github.com/esadr/.
Recommended Citation
Sadrfaridpour, Ehsan, "Fast Machine Learning Algorithms for Massive Datasets with Applications in the Biomedical Domain" (2020). All Dissertations. 2684.
https://tigerprints.clemson.edu/all_dissertations/2684