IRS13 Correlation Measures chapter K. Reitsma J. Sagalyn Harvard University Gerard Salton Use, reproduction, or publication, in whole or in part, is permitted for any purpose of the United States Government. rv-18 for every coefficient to be evaluated. The final evaluation is based on the comparison of these average graphs. The coefficient which produces an average recall-precision graph above and to the right of all the other graphs is assumed to be the best coefficient, with respect to that document collection and that set of queries, since for any value of recall, the precision is higher than for any other coefficient and for any value of precision, the recall is higher than for any other coefficient. It is possible that two average recall-precision graphs may coincide or intersect. In the former case, no conclusion can be made as to which coefficient is better. In the latter case, one coefficient may be better.than another for a given range of recall. For example, is the coeffi- cient D gives an average recall-precision graph above that for coeffi- 1 cient D2 in the range of recall from 0 to .[OCRerr]0 , it may be concluded that coefficient D1 is better than D when the user is interested in the 2 first [OCRerr]O% or less relevant documents. If, however, the user is interested in finding 50%, 75%, or 100% of the relevant documents, D is no longer 1 the most powerful coefficient. The best performance, in this, might result from the use of D to find the first 40% relevant documents and then the 1 use of D to find the remaining relevant documents. 2 One danger exists in using two different coefficients to process a request. It may happen that the specific documents contained in the first 40% retrieved by D are the same documents which D2 retrieves last. In 1 that case, using the two together might not give the desired performance. Most probably, coefficient D1 would be used if the user were only interested in 40% or less of the relevant documents. If he were interested in more, D2 might be used.