Domain knowledge guides the choice of feature combinations whose joint values could be considered noteworthy. In outlier detection, it helps identify useful feature subspaces rather than treating every combination as equally informative. section 10.1.2
This approach prioritizes meaningful interpretation, but it must balance coverage against computation: using more subspaces can improve accuracy, while exhaustive combinations quickly become intractable. section 10.1.2
Useful domain-informed subspaces can be chosen by:
- Selecting features whose combinations have a meaningful interpretation. section 10.1.2
- Combining domain judgment with associations, sparse regions, known outliers, or randomized selection. section 10.1.2
- Using enough features per detector for useful coverage without making detectors computationally excessive. section 10.1.2
