Approximate Query Answering In Numerical Databases

Scienti c databases are usually large, distributed and dynamically changing. We address the problem of e cient processing of queries in scienti c databases, especially in very large numerical databases. Previous work has focused on how to store the database and the design of index structures for the e cient access of data. Recently more and more statistical methods have been used in query optimization. Those methods essentially attempt to approximate the distribution of the attribute values in order to estimate the selectivity of query results. We introduce a new methodology that uses regression techniques to approximate the actual attribute values. Through analysis of the data, one derives a set of characteristic functions to form a \regression database," a compressed image of the original database. Based on these functions, approximate answers to queries may be provided within a pre-speci ed tolerable error, but without the expensive search overhead usually inherent with the use of indexing techniques. We propose a framework to build regression databases. An experimental prototype is implemented to evaluate the technique in terms of realizability, e ciency and practicality. The results demonstrate that our approach is complementary to conventional approaches and to statistical methods.