# Detection of data drift and outliers affecting machine learning model performance over time

arXiv: https://arxiv.org/abs/2012.09258 · HTML: https://mlsanity.com/paper/2012.09258/

## Key facts

- Cited by 19 papers in the map (data through 7 Oct 2026).
- First posted on arXiv 16 Dec 2020.

## Built on by (most cited first, 19 of 19)

- [Holding AI to Account: Challenges for the Delivery of Trustworthy AI in Healthcare](https://mlsanity.com/paper/2211.16444/) (2022)
- [Machine Learning Model Drift Detection Via Weak Data Slices](https://mlsanity.com/paper/2108.05319/) (2021)
- [Systematic Literature Review on Application of Learning-based Approaches in Continuous Integration](https://mlsanity.com/paper/2406.19765/) (2024)
- [Test & Evaluation Best Practices for Machine Learning-Enabled Systems](https://mlsanity.com/paper/2310.06800/) (2023)
- [Systematic Literature Review on Application of Machine Learning in Continuous Integration](https://mlsanity.com/paper/2305.12695/) (2023)
- [Unsupervised Model Drift Estimation with Batch Normalization Statistics for Dataset Shift Detection and Model Selection](https://mlsanity.com/paper/2107.00191/) (2021)
- [Detecting model drift using polynomial relations](https://mlsanity.com/paper/2110.12506/) (2021)
- [Reliable and Interpretable Drift Detection in Streams of Short Texts](https://mlsanity.com/paper/2305.17750/) (2023)
- [Hierarchical Bias-Driven Stratification for Interpretable Causal Effect Estimation](https://mlsanity.com/paper/2401.17737/) (2024)
- [Data Quality Aware Approaches for Addressing Model Drift of Semantic Segmentation Models](https://mlsanity.com/paper/2402.07258/) (2024)
- [Using sequential drift detection to test the API economy](https://mlsanity.com/paper/2111.05136/) (2021)
- [Levels of Autonomous Radiology](https://mlsanity.com/paper/2112.07286/) (2021)
- [Challenges and Solutions to Build a Data Pipeline to Identify Anomalies in Enterprise System Performance](https://mlsanity.com/paper/2112.08940/) (2021)
- [Theory and Practice of Quality Assurance for Machine Learning Systems An Experiment Driven Approach](https://mlsanity.com/paper/2201.00355/) (2022)
- [Generating synthetic multi-dimensional molecular-mediator time series data for artificial intelligence-based disease trajectory forecasting and drug development digital twins: Considerations](https://mlsanity.com/paper/2303.09056/) (2023)
- [A Representation Learning Approach to Feature Drift Detection in Wireless Networks](https://mlsanity.com/paper/2505.10325/) (2025)
- [GTS_Forecaster: a novel deep learning based geodetic time series forecasting toolbox with python](https://mlsanity.com/paper/2509.10560/) (2025)
- [Cloud-Native Evaluation-as-a-Service: A Microservices Architecture for Scalable AI Monitoring with Conformal Guarantees](https://mlsanity.com/paper/2607.21623/) (2026)
- [Testing and Evaluation of Agentic AI Systems In Military Command and Control](https://mlsanity.com/paper/2608.20597/) (2026)

## Abstract (arXiv)

A trained ML model is deployed on another `test' dataset where target feature values (labels) are unknown. Drift is distribution change between the training and deployment data, which is concerning if model performance changes. For a cat/dog image classifier, for instance, drift during deployment could be rabbit images (new class) or cat/dog images with changed characteristics (change in distribution). We wish to detect these changes but can't measure accuracy without deployment data labels. We instead detect drift indirectly by nonparametrically testing the distribution of model prediction confidence for changes. This generalizes our method and sidesteps domain-specific feature representation. We address important statistical issues, particularly Type-1 error control in sequential testing, using Change Point Models (CPMs; see Adams and Ross 2012). We also use nonparametric outlier methods to show the user suspicious observations for model diagnosis, since the before/after change confidence distributions overlap significantly. In experiments to demonstrate robustness, we train on a subset of MNIST digit classes, then insert drift (e.g., unseen digit class) in deployment data in various settings (gradual/sudden changes in the drift proportion). A novel loss function is introduced to compare the performance (detection delay, Type-1 and 2 errors) of a drift detector under different levels of drift class contamination.