Tight coupling of data mining and database systems is a key issue in inductive databases. It ensures scalability, direct and uniform access to both data and patterns stored in databases, as well as proper exploitation of information embedded in the database schema to drive the mining process. In this paper we present a new data mining system, named Mr-SMOTI, which is able to mine (multi-)relational model trees from a tightly coupled relational database. The induced model tree is a (multi-)relational pattern that can be represented by means of a set of selection graphs, which are translated into SQL expressions and stored in XML format. A peculiarity of induced model trees is that they can represent both local and global effects of variables used in regression models. This distinction between local patterns and global models addresses a limitation of current inductive database perspective, which mainly focus on local pattern mining tasks. Preliminary experiments demonstrate the ability of Mr-SMOTI to mine accurate regression predictors from data stored in multiple tables of a relational database
MR-SMOTI: A Data Mining System for Regression Tasks Tightly-Coupled with a Relational Database
APPICE, ANNALISA;CECI, MICHELANGELO;MALERBA, Donato
2003-01-01
Abstract
Tight coupling of data mining and database systems is a key issue in inductive databases. It ensures scalability, direct and uniform access to both data and patterns stored in databases, as well as proper exploitation of information embedded in the database schema to drive the mining process. In this paper we present a new data mining system, named Mr-SMOTI, which is able to mine (multi-)relational model trees from a tightly coupled relational database. The induced model tree is a (multi-)relational pattern that can be represented by means of a set of selection graphs, which are translated into SQL expressions and stored in XML format. A peculiarity of induced model trees is that they can represent both local and global effects of variables used in regression models. This distinction between local patterns and global models addresses a limitation of current inductive database perspective, which mainly focus on local pattern mining tasks. Preliminary experiments demonstrate the ability of Mr-SMOTI to mine accurate regression predictors from data stored in multiple tables of a relational databaseI documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
