Author(s)

Gregory Ditzler

Date Approved

7-9-2011

Document Type

Thesis

Degree Name

M.S. Engineering

Department

Electrical and Computer Engineering

College

Henry M. Rowan College of Engineering

First Advisor

Polikar, Robi

Subject(s)

Machine learning

Disciplines

Electrical and Computer Engineering

Abstract

Learning data sampled from a nonstationary distribution has been shown to be a very challenging problem in machine learning, because the joint probability distribution between the data and classes evolve over time. Thus learners must adapt their knowledge base, including their structure or parameters, to remain as strong predictors. This phenomenon of learning from an evolving data source is akin to learning how to play a game while the rules of the game are changed, and it is traditionally referred to as learning concept drift. Climate data, financial data, epidemiological data, spam detection are examples of applications that give rise to concept drift problems. An additional challenge arises when the classes to be learned are not represented (approximately) equally in the training data, as most machine learning algorithms work well only when the class distributions are balanced. However, rare categories are commonly faced in real-world applications, which leads to skewed or imbalanced datasets. Fraud detection, rare disease diagnosis, anomaly detection are examples of applications that feature imbalanced datasets, where data from category are severely underrepresented. Concept drift and class imbalance are traditionally addressed separately in machine learning, yet data streams can experience both phenomena. This work introduces Learn++.NIE (nonstationary & imbalanced environments) and Learn++.CDS (concept drift with SMOTE) as two new members of the Learn++ family of incremental learning algorithms that explicitly and simultaneously address the aforementioned phenomena. The former addresses concept drift and class imbalance through modified bagging-based sampling and replacing a class independent error weighting mechanism - which normally favors majority class - with a set of measures that emphasize good predictive accuracy on all classes. The latter integrates Learn++.NSE, an algorithm for concept drift, with the synthetic sampling method known as SMOTE, to cope with class imbalance. This research also includes a thorough evaluation of Learn++.CDS and Learn++.NIE on several real and synthetic datasets and on several figures of merit, showing that both algorithms are able to learn in some of the most difficult learning environments.

Share

COinS