Selectivity Estimation in Extensible Databases - A Neural Network Approach

Abstract

Extensible database systems allow users to create new data types representing spatial, text, image, and other multimedia objects, and functions which operate on the new types[S96]. The new data types and functions can be used as predicates in SQL queries. Consider an insurance industry application, built using an extensible DBMS that has user defined extensions for spatial, text, date/time, and other business objects. The extensions support new data types such as point, circle, etc., and functions such as confains, interval, text-contains, etc. Let the tables Policies (policy-id, name, address, location, vehicle-type, . . .) and Claims (policy-id, claim-tag, accident-date, accident-location, accident-report, . . .) represent the partial schema containing both SQL’92 and user defined data types (UDTs). Consider a scenario in a targeted marketing application that requires a mailing list of all customers within 5 miles of point L, who have insured a ‘sports utility vehicle’ and were involved in a ‘rear-ended’ accident in the past 3 years. The corresponding SQL query would be: SELECT P.name, P.address FROM Policies P, Claims C WHERE coniuins(P.location, circle(L, 5)) AND P.vehicle type=“Sports Utility” AND P.policy-id = C.policy-id AND texl_contuins(C.accident-report, “rear-ended”) AND inietval(C.accident-date,current_date)<3 years