Proc. of 4th IEEE Int'l Conf. on Parall. and Distr. Inf. Syst., pp 220 - 31,
Dec. 18 - 20, 1996
------------------------------

  
        Building Regression Cost Models for Multidatabase Systems

                               Qiang Zhu
             Department of Computer and Information Science
                  The University of Michigan - Dearborn
                          Dearborn, MI 48128, USA
 
                             Per-Ake Larson*
                    Department of Computer Science
                         University of Waterloo
                   Waterloo, Ontario N2L 3G1, Canada
               * Current address: Microsoft Corporation
                          One Microsoft Way
                     Redmond, WA 98052-6399, USA
 
 
                               ABSTRACT

A major challenge for performing global query optimization in a 
multidatabase system (MDBS) is the lack of cost models for local database
systems at the global level. In this paper we present a statistical 
procedure based on multiple regression analysis for building cost models 
for local database systems in an MDBS. Explanatory variables that can be
included in a regression model are identified and a mixed forward and 
backward method for selecting significant explanatory variables is
presented. Measures for developing useful regression cost models, such as
removing outliers, eliminating multicollinearity, validating regression
model assumptions, and checking significance of regression models, are
discussed. Experimental results demonstrate that the presented
statistical procedure can develop useful local cost models in an MDBS. 


KEYWORDS:  multidatabase system, global query optimization, cost model, 
           cost estimation, multiple regression