Statistical Approaches to Predictive Modeling

Prediction, i.e., predicting the potential values or value distributions of certain attributes for objects in a database or data warehouse, is an attractive goal in data mining. To predict future events not shown in databases with high quality can help users to make smart business decisions. With the concern of both scalability and high quality of prediction, we propose a predictive modeling algorithm for interactive prediction in large databases and data warehouses. The algorithm consists of three steps: (1) data generalization, which converts data in relational databases or data warehouses into a multi-dimensional databases to which e cient analysis techniques can be applied; (2) relevance analysis, which identi es the attributes that are highly relevant to the prediction, to reduce number of attributes in prediction with the bene ts in improving both e ciency and reliability of prediction; and (3) a statistical regression model, called generalized linear model, is constructed for high quality prediction. We explore two types of model featuring di erent problems. Moreover, with this method, a user can interact with a data mining system by presenting probes with constants at di erent levels of abstraction and attempt to predict values of a predicted attribute at di erent levels of abstraction. Also, a user may drill-down or roll-up along any attribute dimensions and then do prediction analysis. Our analysis and experimental results show that the method provides high prediction quality with modest or intermediate data generalization and it leads to e cient, interactive prediction in large databases. iii Acknowledgments I am deeply grateful to my senior supervisor, Dr. Jiawei Han, for all assistance given to me since I enrolled in this graduate program, especially for his invaluable advice, guidance and the time that he spent with me in the preparation of this thesis. I wish to express my gratitude to my committeemembers, Dr. Veronica Dahl and Dr. Qiang Yang for their valuable and insightful comments and suggestions. I would like to extend my thanks to all other faculty members of the School of Computing Science for all the courses I have taken from them. My thanks also go to fellow graduate students, especially those who have been working in the Intelligent Database System Lab, for the help and friendship they have given to me. I shall particularly mention Wan Gong, Jenny Chiang, Sonny Chee, and Shuhua Zhang for their great help through my research. I especially thank the School of Computing Science at Simon Fraser University for providing me the opportunity to study here. My sincere thanks are given to Mrs. E. Krbavac, Mrs. K. Jaager, Mrs. C. Edwards, and Mrs. W. Davis for their assistance. iv Dedication To my parents v