DETERMINING INFORMATION-RETRIEVAL AND FILTERING PERFORMANCE WITHOUT EXPERIMENTATION

The performance of an information retrieval or text and media filterin g system may be determined through analytic methods as well as by trad itional simulation or experimental methods. These analytic methods can provide precise statements about expected performance. They can thus determine which of two similarly performing systems is superior. For b oth a single query term and for a multiple query term retrieval model, a method for comparing the performance of different probabilistic ret rieval methods is developed. This method may be used in computing the average search length for a query, given only knowledge of database pa rameter values. Predictive models for inverse document frequency, bina ry independence, and relevance feedback based retrieval and filtering are described. Simulations illustrate how the single term model perfor ms and sample performance predictions are given for single term and mu ltiple term problems.