Interactivity, Scalability and Resource Control for Efficient KDD Support in DBMS

Abstract

The conflict between resource consumption and query performance in the data mining context often has no satisfactory solution. This is in sharp contrast to the needs of the analysts for interactive response times and has rendered the seamless integration of data mining operators into common multiuser database systems a difficult and (so far) not very successful task. This paper describes an approach that allows to combine preprocessing and data mining operators into one common KDD-aware implementation algebra such that interactivity, scalability and resource efficiency can simultaneously be achieved. The basic idea of our framework is pipelining. However, since there is a danger of blocking pipelines, we introduce controlled ordering-, cardinality- and special-value-properties of the data stream across the whole query tree up to the complex data mining operators. The framework builds on a spezialized index that is basically an extension of the UB-Tree and efficiently provides various data orderings. These orderings and the remaining properties are then exploited by the KDD-algebra operators to release results and internal data structures early enough to allow pipelined, resource-efficient query processing with interactive response times. This paper describes the framework and demonstrates its benefits in preprocessing and in the parallel and interactive detection of outliers.