DDMDriftDetector<T>
Models & Types · AiDotNet.DriftDetection
Drift Detection Method (DDM) for concept drift detection based on error rate monitoring.
Inheritance: DriftDetectorBase<T> → DDMDriftDetector<T>
For Beginners
DDM monitors the error rate of a classifier over time. When the error rate increases significantly beyond what was observed during a stable period, drift is detected. DDM is simple, fast, and effective for detecting sudden drift.
How It Works
How it works:
- Track the error rate p and its standard deviation s = sqrt(p(1-p)/n)
- Remember the minimum observed value of p + s (the baseline)
- Warning level: p + s > p_min + s_min + warning_threshold × s_min
- Drift level: p + s > p_min + s_min + drift_threshold × s_min
Key insight: For a stable distribution, p + s should remain relatively constant. A significant increase indicates the underlying distribution has changed.
Parameters:
- warningThreshold: Number of standard deviations for warning (default: 2)
- driftThreshold: Number of standard deviations for drift (default: 3)
Advantages:
- Simple and computationally efficient
- Two-stage detection (warning before drift)
- Well-suited for sudden drift
Limitations:
- Assumes binary errors (0 or 1)
- May be slow to detect gradual drift
- Can miss drift if error rate decreases
Reference: Gama et al. (2004). “Learning with Drift Detection”
Constructors
| Constructor | Summary |
|---|---|
DDMDriftDetector(double warningThreshold, double driftThreshold, int minimumObservations) |
Creates a new DDM drift detector. |
Properties
| Property | Summary |
|---|---|
bool IsInWarning { get; set; } |
Gets whether the detector is in the warning zone. |
double ErrorRate { get; } |
Gets the current error rate. |
double MinimumPPlusS { get; } |
Gets the minimum p + s observed. |
int SamplesInWarning { get; } |
Gets the number of samples since entering warning zone. |
Methods
| Method | Summary |
|---|---|
bool AddObservation(T value) |
Adds a new observation (typically a prediction error: 1 for wrong, 0 for correct). |
void Reset() |
Resets the detector to its initial state. |