09-19, 14:40–15:15 (Europe/Amsterdam), Van Gogh
This talk will delve into the technical and conceptual challenges associated with drift detection on irregular time series exhibiting non-uniform seasonal patterns, such as end-of-month, pay-day, or holiday effects. We will demonstrate how drifts can be efficiently identified using a combination of the MIST (Multiple Irregular Seasonalities and Trend decomposition) and DTW (Dynamic Time Warping) algorithms.
Persistent, long-term drifts in time series pose a significant challenge as they can evade detection over extended periods, potentially lasting for months or even years. To ensure the universal applicability of the drift detection system across a wide range of domains, it is crucial that the system is equipped to manage missing values or, more broadly, irregular time series. This functionality would allow the end users of the system to apply it directly to the raw time series data, requiring minimal or no data pre-processing.
This talk will introduce you to the various challenges involved in monitoring drifts across large-scale datasets comprising millions of time series. Issues such as irregular time points, missing data, uneven weighting of observations, and complex seasonal patterns will be highlighted. To address these challenges, we propose an efficient four-step drift detection procedure:
- De-seasonalization with MIST: Utilizing the MIST (Multiple Irregular Seasonalities and Trend decomposition) model, which adapts the MSTL algorithm [1] to handle irregular time series and non-uniform seasonality patterns.
- Drift segmentation with DTW: Use Dynamic Time Warping to segment the time series into reference, dip and recovery series.
- Statistical drift assessment: Determining whether a drift is present based on statistical criteria.
- Alert promotion: Deciding whether the detected drift warrants an alert based on various business rules.
Outline
- Minutes 0-5: Problem Statement
- Objectives and overview of the four-step drift detection procedure.
- Minutes 5-15: MIST Model
- Explanation of the MIST model for de-seasonalizing time series and comparison with existing off-the-shelf solutions
- Minutes 15-20: DTW Segmentation
- Use of DTW for segmenting time series into reference, dip, and recovery segments.
- Minutes 20-25: Drift Assessment and Alert Promotion - Practical considerations for drift detection and criteria for alerting.
- Minutes 25-30: Limitations and Future Work
Prior Knowledge
No prior knowledge is required but familiarity with classic seasonality-trend decomposition techniques and DTW algorithm could be helpful.
References
[1] Bandara, Kasun, Rob J. Hyndman, and Christoph Bergmeir. "MSTL: A seasonal-trend decomposition algorithm for time series with multiple seasonal patterns." arXiv preprint arXiv:2107.13462 (2021).
As a senior machine learning scientist at Adyen, my current focus lies in the development of models and explainability tools to monitor and comprehend the intricacies of payment processes. My primary areas of interest encompass Bayesian probabilistic modeling, machine learning explainability, and causality. In my past roles as a freelance professional, employed data scientist, and university researcher, I have engaged with a wide array of applied modeling challenges, such as churn modeling, matching engines, anomaly detection, recommender systems, and modeling of human behavior under risk.