Vitalie Spinu - Drift Detection on Irregular Time Series with Multiple Non-Uniform Seasonal Patterns

Drift detection system handles millions of irregular time series data points focusing on payment flows and authorization rates at ADIEN

Algorithm adapts to complex seasonality patterns including hourly, daily, weekly, and monthly cycles while handling irregular intervals and missing data

Four-step procedure:

• Remove predictable patterns
• Segment time series data
• Perform statistical testing
• Make business decisions on alerts

System incorporates volume weighting to prioritize significant drops and reduce false alerts

DTW (Dynamic Time Warping) algorithm modified to work with irregular time series and handle reference/dip period detection

Global optimization approach trades off between recency and magnitude of drops when detecting anomalies

Algorithm can process data in milliseconds, making it suitable for online/real-time monitoring

System reduces redundancy by smart grouping of related alerts across payment methods (e.g., Visa/Mastercard)

Handles multiple types of drifts:

• Incremental slow-moving drifts
• Sudden drops
• Recovery patterns
• Seasonal variations

Framework allows for plug-and-play components at each step while maintaining weight integration throughout the process