Hello, BobDuin.
I encounter a problem. Here is a dataset with two classes, target and false target. When I classify it with bpxnc or adaboostc, the errors of target are always high, and the errors of false target are quit low. Certainly flase target is even and homogeneous, and target are not. But how can I interpet it with Pattern recognition theory? 
Thanks a lot.

Understanding the configuration of high-dimensional dataset takes a lot of experimentation. Results also depend on class sizes and priors. You may consider to use one-class classifiers. Have look at the web-site of David Tax.

Best regards,

Bob Duin