Robust Z-Score#
A statistical anomaly detector that uses modified Z-Scores that are robust to preexisting outliers in the training set. The modified Z-Score is defined as the feature value minus the median over the median absolute deviation (MAD). Anomalies are flagged if their final weighted Z-Score exceeds a user-defined threshold.
Note
A beta value of 1 means the estimator only considers the maximum absolute Z-Score, whereas a setting of 0 indicates that only the average Z-Score factors into the final score.
Interfaces: Estimator, Learner, Scoring, Persistable
Data Type Compatibility: Continuous
Parameters#
# | Name | Default | Type | Description |
---|---|---|---|---|
1 | threshold | 3.5 | float | The minimum Z-Score to be flagged as an anomaly. |
2 | beta | 0.5 | float | The weight of the maximum Z-Score in the overall anomaly score. |
3 | smoothing | 1e-9 | float | The amount of epsilon smoothing added to the MAD of each feature. |
Example#
use Rubix\ML\AnomalyDetectors\RobustZScore;
$estimator = new RobustZScore(3.5, 0.25, 1e-6);
Additional Methods#
Return the median of each feature column in the training set:
public medians() : float[]|null
Return the median absolute deviation (MAD) of each feature column in the training set:
public mads() : float[]|null
References#
-
B. Iglewicz et al. (1993). How to Detect and Handle Outliers. ↩