<title>Using boundary methods as ordering indices on ATR feature sets</title>

<title>Theoretical and complexity issues for feature set evaluation using boundary methods</title>

“…Past justification of 05 being a viable measure of class separability has been based upon empirical evidence.11 Figure 6 shows evidence of a nearly linear relationship between Bayes error and a modified 05. Details of these early experiments and the the definition of the modified 05 can be found in [11]. Demonstrating that a formal relationship between 05 and Bayes error exists would give a theoretical justification for the use of 05 for FSE, since this would give a connection between the 05 and a measure of class separability known to be optimal.'…”

“…Specifically: 0 r(jto) = akJ.s(kto) + c k=j (11) where k,j is the percentage the kth differential BM, overlap sum s(kt.u) contributes to the jth differential reject rate r(jto). For example, from Figure 8, a,n_i,m is equal to the ratio of area 2 to the sum of areas 2 and 4.…”

“…This paper investigates the ability of BMs to obtain a new class separability measure. The use of BMs as a FSE technique was first introduced in [11]. This paper investigates the relationship between BMs measure of class separability, known as the overlap sum (OS) , to Bayes error.…”