A categorical outlier detector analyzes features whose values belong to discrete categories. It is useful when categorical data dominates and avoids the difficulties of creating numeric encodings, but relatively few such detectors exist. section 3.2.4
Frequent item set methods identify rows containing unusually few or infrequent combinations of values. Frequent pattern outlier factor (FPOF) scores a row from the frequent item sets it contains: inliers generally contain more and more frequent sets than outliers. Association rules provide another frequent-item-set approach. section 3.2.4
Apply categorical detection with these practical decisions:
- Convert numeric and date columns into categories through binning when using a primarily categorical method; choose bins with interpretability and trade-offs in mind. section 3.2.4 section 3.3.2
- For mixed tables, either numerically encode categorical columns for a numeric detector or bin numeric columns for a categorical detector; few detectors support both types directly. section 3.3.2
- Interpret scores as evidence for inspection rather than automatic meaning: evaluation is difficult, often requiring manual review, representative data, and ongoing tuning. section 4.15
- Entropy-based detectors score individual categorical cell values and combine them per row, without considering associations between features; they therefore emphasize univariate rarity rather than combinations. section 7.9
Expense and preference records
In staff expenses, repeated Staff ID–Account combinations form frequent patterns; a row containing substantially fewer or less frequent combinations can receive a higher outlier score. In a SpeedDating dataset, FPOF uses one-hot categorical values and a minimum-support threshold to find frequent item sets. Its most outlier row combines unusual preference ranges, while the most inlier row contains common combinations.
