Data annotation assigns labels or target values to records. In outlier detection, hand-labeling evaluates whether flagged records matter to the project, rather than merely asking whether they are statistically unusual. Domain experts and data scientists may assign binary relevance labels, specific outlier categories, and importance scores. section 15.1
Useful annotation practice includes:
- Define specific, consistently applied categories for recurring or important outliers; a record may receive multiple labels. section 15.1.1
- Record statistically unusual but unimportant cases separately, so models can deemphasize or exclude them from future investigation. section 15.1.1
- Use labeled records to retune detectors, track outlier types, group investigations, filter known cases, and create rules, classifiers, or regression tests. section 15.6
Financial and sales records
A financial transaction entered by unusually senior staff might receive an “unusual or inappropriate staff for transaction” label, even if investigation finds no immediate issue. Sales records involving rare combinations of items and time of day might receive a “rare time of day for item sold” label when the pattern is unusual but not worth investigating. section 15.1 section 15.1.1
Hand-labeled project data can also support detector comparisons through doping experiments, although classification datasets provide only a general starting point. The most suitable detector depends on the current dataset and the kinds of outliers that matter. section 8.6
