CCE Theses and Dissertations

Date of Award


Document Type


Degree Name

Doctor of Philosophy (PhD)


Graduate School of Computer and Information Sciences


Junping Sun

Committee Member

Michael J. Laszlo

Committee Member

Lee Leitner


Databases are growing exponentially in many application domains. Timely construction of models that represent identified patterns and regularities in the data facilitate the prediction of future events based upon past performance. Data mining can promote this process through various model building techniques. The goal is to create models that intuitively represent the data and perhaps aid in the discovery of new knowledge.

Most data mining methods rely upon either fully-automated information-theoretic or statistical algorithms. Typically, these algorithms are non-interactive, hide the model derivation process from the user, require the assistance of a domain expert, are application-specific, and may not clearly translate detected relationships.

This paper proposes a visual data mining algorithm, BLUE, as an alternative to present data mining techniques. BLUE visually supports the processes of classification and prediction by combining two visualization methods. The first consists of a modification to independence diagrams, called BIDS, allowing for the examination of pairs of categorical attributes in relational databases. The second uses decision trees to provide a global context from which a model can be constructed. Classification rules are extracted from the decision trees to assist in concept representation. BLUE uses the abilities of the human visual system to detect patterns and regularities in images. The algorithm employs a mechanism that permits the user to interactively backtrack to previously visited nodes to guide and explore the creation of the model. As a decision tree is induced, classification rules are simultaneously extracted. Experimental results show that BLUE produces models that are more comprehensible when compared with alternative methods. These experimental results lend support for future studies in visual data mining.

  Link to NovaCat