EDDMDriftDetector<T>
Models & Types · AiDotNet.DriftDetection
Early Drift Detection Method (EDDM) for concept drift detection.
Inheritance: DriftDetectorBase<T> → EDDMDriftDetector<T>
For Beginners
EDDM is an improvement over DDM that monitors the distance between errors rather than just the error rate. This makes it better at detecting gradual drift because even if the error rate stays similar, the spacing between errors may change.
How It Works
How it works:
- Track the distance (number of samples) between consecutive errors
- Calculate the mean distance p’ and standard deviation s’
- Monitor p’ + 2s’ and remember its maximum value
- Warning: when (p’ + 2s’) / (p’_max + 2s’_max) drops below warning threshold
- Drift: when (p’ + 2s’) / (p’_max + 2s’_max) drops below drift threshold
Key insight: When performance is good, errors are far apart (high mean distance). When drift occurs, errors become more frequent and closer together (lower mean distance).
Advantages over DDM:
- Better at detecting gradual drift
- More sensitive to changes in error patterns
- Can detect drift even when error rate is low
Parameters:
- warningThreshold: Ratio for warning level (default: 0.95)
- driftThreshold: Ratio for drift level (default: 0.90)
Reference: Baena-García et al. (2006). “Early Drift Detection Method”
Constructors
| Constructor | Summary |
|---|---|
EDDMDriftDetector(double warningThreshold, double driftThreshold, int minimumObservations, int minimumErrors) |
Creates a new EDDM drift detector. |
Properties
| Property | Summary |
|---|---|
double AverageErrorDistance { get; } |
Gets the average distance between errors. |
double CurrentRatio { get; } |
Gets the current ratio (lower means closer to drift). |
double MaximumPPrime2S { get; } |
Gets the maximum p’ + 2s’ observed. |
int ErrorCount { get; } |
Gets the number of errors detected. |
int MinimumErrors { get; } |
Gets the minimum number of errors required before detection starts. |
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. |