Date of Award

Fall 2019

Document Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

Computational Analysis and Modeling

First Advisor

Sumeet Dua

Abstract

A non-stationary dataset is one whose statistical properties such as the mean, variance, correlation, probability distribution, etc. change over a specific interval of time. On the contrary, a stationary dataset is one whose statistical properties remain constant over time. Apart from the volatile statistical properties, non-stationary data poses other challenges such as time and memory management due to the limitation of computational resources mostly caused by the recent advancements in data collection technologies which generate a variety of data at an alarming pace and volume. Additionally, when the collected data is complex, managing data complexity, emerging from its dimensionality and heterogeneity, can pose another challenge for effective computational learning. The problem is to enable accurate and efficient learning from non-stationary data in a continuous fashion over time while facing and managing the critical challenges of time, memory, concept change, and complexity simultaneously.

Feature space modeling is one of the most effective solutions to address this problem. For non-stationary data, selecting relevant features is even more critical than stationary data due to the reduction of feature dimension which can ensure the best use a computational resource to produce higher accuracy and efficiency by data mining algorithms. In this dissertation, we investigated a variety of feature space modeling techniques to improve the overall performance of data mining algorithms. In particular, we built Relief based feature sub selection method in combination with data complexity iv analysis to improve the classification performance using ovarian cancer image data collected in a non-stationary batch mode. We also collected time series health sensor data in a streaming environment and deployed feature space transformation using Singular Value Decomposition (SVD). This led to reduced dimensionality of feature space resulting in better accuracy and efficiency produced by Density Ration Estimation Method in identifying potential change points in data over time. We have also built an unsupervised feature space modeling using matrix factorization and Lasso Regression which was successfully deployed in conjugate with Relative Density Ratio Estimation to address the botnet attacks in a non-stationary environment.

Relief based feature model improved 16% accuracy of Fuzzy Forest classifier. For change detection framework, we observed 9% improvement in accuracy for PCA feature transformation. Due to the unsupervised feature selection model, for 2% and 5% malicious traffic ratio, the proposed botnet detection framework exhibited average 20% better accuracy than One Class Support Vector Machine (OSVM) and average 25% better accuracy than Autoencoder. All these results successfully demonstrate the effectives of these feature space models.

The fundamental theme that repeats itself in this dissertation is about modeling efficient feature space to improve both accuracy and efficiency of selected data mining models. Every contribution in this dissertation has been subsequently and successfully employed to capitalize on those advantages to solve real-world problems. Our work bridges the concepts from multiple disciplines ineffective and surprising ways, leading to new insights, new frameworks, and ultimately to a cross-production of diverse fields like mathematics, statistics, and data mining.

Share

COinS