Date of Award

Spring 2010

Document Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

Computational Analysis and Modeling

First Advisor

Sumeet Dua

Abstract

The Internet era has revolutionized computational sciences and automated data collection techniques, made large amounts of previously inaccessible data available and, consequently, broadened the scope of exploratory computing research. As a result, data mining, which is still an emerging field of research, has gained importance because of its ability to analyze and discover previously unknown, hidden, and useful knowledge from these large amounts of data. One aspect of data mining, known as frequent pattern mining, has recently gained importance due to its ability to find associative relationships among the parts of data, thereby aiding a type of supervised learning known as "associative learning".

The purpose of this dissertation is two-fold: to develop and demonstrate supervised associative learning in non-temporal data for multi-class classification and to develop a new frequent pattern mining algorithm for time varying (temporal) data which alleviates the current issues in analyzing this data for knowledge discovery. In order to use associative relationships for classification, we have to algorithmically learn their discriminatory power. While it is well known that multiple sets of features work better for classification, we claim that the isomorphic relationships among the features work even better and, therefore, can be used as higher order features. To validate this claim, we exploit these relationships as input features for classification instead of using the underlying raw features. The next part of this dissertation focuses on building a new classifier using associative relationships as a basis for the multi-class classification problem. Most of the existing associative classifiers represent the instances from a class in a row-based format wherein one row represents features of one instance and extract association rules from the entire dataset. The rules formed in this way are known as "class constrained rules," as they have class labels on the right side of the rules. We argue that this class constrained representation schema lacks important information that is necessary for multi-class classification. Further, most existing works use either the intraclass or inter-class importance of the association rules, both of which sets of techniques offer empirical benefits. We hypothesize that both intra-class and inter-class variations are important for fast and accurate multi-class classification. We also present a novel weighted association rule-based classification mechanism that uses frequent relationships among raw features from an instance as the basis for classifying the instance into one of the many classes. The relationships are weighted according to both their intra-class and inter-class importance.

The final part of this dissertation concentrates on mining time varying data. This problem is known as "inter-transaction association rule mining" in the data-mining field. Most of the existing work transforms the time varying data into a static format and then use multiple scans over the new data to extract patterns. We present a unique index-based algorithmic framework for inter-transaction association rule mining. Our proposed technique requires only one scan of the original database. Further, the proposed technique can also provide the location information of each extracted pattern. We use mathematical induction to prove that the new representation scheme captures all underlying frequent relationships.

Share

COinS