I have a commercial outlier detection problem in moderate dimension (8-25).
We have a limited number of true positive tags and can roughly evaluate performance of various methods. So far, the 1-class SVM has beat the competitors I've tried.
My management is uncomfortable though with the scoring performance and representation of trained SVM's as substantial number of support vectors need to be kept. Relevance Vector Machines in classification and regression offer competitive performance with far fewer stored vectors necessary, but I have never seen a 1-class RVM algorithm description or implementation in my brief searches.
Question: does anybody know of a paper about 1-class RVM or something reasonably similar with good outlier performance? Does anybody know of an implementation? (Python/java is ideal).