OA: Generic cluster-based drift detection

[detlefarend]

Description/Motivation
Cluster-based drift detection is based on observing changes in single or multiple cluster properties. Similar to cluster-based anomaly detection, observing single or multiple properties in a generic way is imaginable. Relevant parameters for such an algorithm could be

• Cluster property/ies to be observed
• Threshold/s for each property (lower/upper threshold -> hysteresis to avoid oscillation)
• Drift type to be raised
• Visualization: color for drifting clusters

Details:

• Stages of change: gradient of 1., 2., ... n. order of the property over the time ("velocity", "acceleration", ...)
• Strategies for observing multiple properties
• Buffering of drifts for tracking
• Detection of starting or ending drifts

Task list

• 1. Conception and design
  • Stage 1: single property observation
  • Stage 2: multiple property observation
• 2. Implementation (Template: mlpro.oa.streams.tasks.driftdetection.clusterbased.DriftDetectorCB)
• 3. Validation
• 4. Demonstration
• 5. Documentation

Related issues
#1138

Cross references
...