Drift Metrics¶
Data or influence drift metrics compare a feature or influence values across two different data splits. TruEra uses 'range' to indicate a numerical feature's domain size, and 'card' to indicate the cardinality of a categorical feature's domain. In all cases, a result of 0 indicates no drift.
Name | Feature Type | Range | Notes |
---|---|---|---|
Wasserstein | Numerical | [0, range] | Also referred to as "Earth Mover's Distance". |
Wasserstein (unordered) | Categorical | [0, 1] | Same as Total Variation Distance. |
Wasserstein (ordered) | Categorical | [0, card] | Interprets categorical values as integers arranged on a [0, card] line. |
Total variation distance | Categorical | [0, 1] | Half of L1 Distance. |
Jensen-Shannon distance | Both | [0, 1] | |
Chi-Square test | Categorical | [0, +inf) | |
Population stability index | Both | [0, +inf) | Numeric version uses base deciles as segments. Not symmetric. |
L1 | Categorical | [0, 2] | |
L2 | Categorical | [0, sqrt(2)] | |
LInfinity | Categorical | [0, 1] | |
Difference of mean | Numerical | (-inf, +inf) | Not a metric. Not symmetric. Directional and can have cancellation. |
Energy distance | Numerical | [0, +inf) | |
Kolmogorov-Smirnov statistic | Numerical | [0, +inf) |
Click Next below to continue.