Retraining fits an outlier-detection model again as data, detection goals, or the supply of labeled examples changes. It can use newer data with the existing configuration or introduce substantive modeling changes. section 15.5
A practical retraining cycle includes:
- Collect substantially more labeled data, including records of interest and records that are not. section 15.5
- Choose between training on recent data with the previous setup or changing detectors, features, parameters, distance metrics, preprocessing, or ensemble combination methods. section 15.5
- Run regression tests to verify that known outliers remain highly scored without excessively scoring labeled non-outliers. section 15.5
Checking a retrained detector
After retraining an outlier detector, score a stored list of labeled records. Confirm that previously identified outliers still receive high scores and that records labeled as not interesting do not receive high scores too often. section 15.5
