ORCID
https://orcid.org/0000-0002-2694-3096
Date of Award
Summer 2024
Language
English
Embargo Period
7-11-2026
Document Type
Dissertation
Degree Name
Doctor of Philosophy (PhD)
College/School/Department
Department of Electrical and Computer Engineering
Program
Electrical and Computer Engineering
First Advisor
Daphney-Stavroula Zois
Committee Members
Daphney-Stavroula Zois, Nathan Dahlin, Dola Saha, Charalampos Chelmis
Keywords
supervised classification, ensemble learning, instance-wise feature acquisition, classifier selection, structure learning
Subject Categories
Other Electrical and Computer Engineering
Abstract
Traditional supervised classification typically involves assigning a label to a single variable, considering a common subset of features for all the instances, and employing a single classifier. However, in many real–world applications like behavioral analysis or insurance recommendation, a data instance is described by a set of related variables such as physical activity and emotion, or driving quality and accident severity. At the same time, these variables are not directly observable but can be inferred via noisy but costly features. Specifically, access to all features is prohibitive due to cost, invasiveness, or limited resources. Finally, while using a single classifier may be suitable for some instances, others may require the use of one or more simple or advanced pre–trained models to be correctly predicted. To address the above challenges, this thesis presents mathematical frameworks and algorithms for datum–wise learning and inference in the context of supervised classification.
First, a framework is proposed to classify related and unobservable variables while keeping the feature acquisition cost to a minimum. Specifically, the relationships between variables are modeled by a known Bayesian network structure. The proposed framework dynamically acquires features in a sequential manner and reaches a classification decision for each variable using a subset of the acquired features. Classification decisions are propagated through the network by exploiting the Bayesian network properties. The proposed framework outperforms existing classification and feature acquisition methods in terms of accuracy and the average number of acquired features. Unfortunately, in many cases, the structure of the Bayesian network is unknown. To address this challenge, a method is proposed that sequentially evaluates variable relationships until it makes a specific decision about the underlying Bayesian network structure. The proposed method speeds–up variable relationship identification without compromising accuracy.
Second, an instance–wise feature acquisition and classifier selection mathematical framework is presented to further enhance the accuracy. The proposed framework sequentially acquires features until it determines that additional features will not improve label assignment. At that stage, easy–to–classify instances are handled by a simple classifier, while difficult–to–classify instances are forwarded to one out of a number of advanced classifiers. This improves overall accuracy while reducing the average number of acquired features. Further, the proposed framework is extended to classify variables related via a known Bayesian network structure.
Finally, inspired by the success of ensemble learning, a mathematical framework that considers the more general problem of determining both the features and the expert decisions to be acquired in a datum–wise fashion is proposed. The framework initially acquires features one by one, until the feature acquisition process terminates. Then it either assigns a label or switches to acquire expert decisions followed by the label assignment. The proposed framework further enhances accuracy compared to existing methods acquiring fewer features and expert decisions on average.
License
This work is licensed under the University at Albany Standard Author Agreement.
Recommended Citation
Ekanayake Mudiyanselage, Sachini Piyoni Ekanayake, "Datum–Wise Learning and Inference for Supervised Classification" (2024). Electronic Theses & Dissertations (2024 - present). 5.
https://scholarsarchive.library.albany.edu/etd/5