Good post on the exact problem I ran into when developing a machine learning algorithm to identify rule violations (basically fraud) based on self-reported desc

Good post on the exact problem I ran into when developing a machine learning algorithm to identify rule violations (basically fraud) based on self-reported descriptive text data. The few records that are "fraud" are precious from a training standpoint. I found it helpful to start by over-representing the rare class and then slowly bring the balance back to the observed percentages. The main reason for my over-representation of the rare class was that rare case was often not correctly identified (many false negatives) in the so-called "Gold Standard" training data. By over-representing the rare case, I was able to build a better training set while tuning your algorithm. I'd be interested to hear if others flush out false positives in a similar/different manner (for the Skewed/Rare Class problems described by the article).