Selecting adequate samples for approximate decision support queries

Abstract

For highly selective queries, a simple random sample of records drawn from a large data warehouse may not contain sufficient number of records that satisfy the query conditions. Efficient sampling schemes for such queries require innovative techniques that can access records that are relevant to each specific query. In drawing the sample, it is advantageous to know what would be an adequate sample size for a given query. This paper proposes methods for picking adequate samples that ensure approximate query results with a desired level of accuracy. A special index based on a structure known as the k-MDI Tree is used to draw samples. An unbiased estimator named inverse simple random sampling without replacement is adapted to estimate adequate sample sizes for queries. The methods are evaluated experimentally on a large real life data set. The results of evaluation show that adequate sample sizes can be determined such that errors in outputs of most queries are wtihin the acceptable limit of 5%.