An interpretable model makes its decisions understandable through meaningful inputs and relationships. In outlier detection, interpretability comes from models whose scoring logic can be inspected directly, rather than from explanations added after fitting. section 13.1.1
Common interpretable approaches include:
- Univariate tests and distribution-based detectors flag unusually small or large values, making the extreme feature values behind a score visible. Examples include interquartile range, interdecile range, median absolute deviation, HBOS, and ECOD. section 13.1.1
- Frequent pattern outlier factor exposes frequent item sets and shows which sets each flagged row contains or lacks. section 13.1.1
- Association rules expose which rules a row respects or violates, including whether the left-hand side and right-hand side are present. section 13.1.1
Interpretability is relative rather than automatic. Many features can make it difficult to compare two rows with several extreme values, and even an interpretable score may not provide a complete or perfectly reliable account of the detector. section 13.1.1 section 13.2.5
