Model Monitoring and Drift Resources

Overview

Curated links on model monitoring, model performance, model drift and A/B testing.
Quotes:
- “Ensuring high model performance in live deployments is arguably the most important aspect of monitoring machine learning systems” - Monitoring and explainability of models in production - 2020-07-13
- “The lifecycle of a machine learning model only begins once it’s in production” - Saucedo - 2020-12-13

MLflow and Model Monitoring

https://stackoverflow.com/questions/65937786/data-and-model-drift-monitoring-with-mlflow - Stack Overflow - 2021-01-28

Vendor Products

Databricks

Databricks blogs

Productionizing Machine Learning: From Deployment to Drift Detection - blog - 2019-09-18
- Webinar slides as PDF
- https://github.com/joelcthomas/modeldrift - code

SAIS/DAIS

Drifting Away: Testing ML Models in Production - slides - Chengyin, Niall - DAIS 2021
- https://github.com/chengyin38/dais_2021_drifting_away
KFServing, Model Monitoring with Apache Spark and a Feature Store - Dowling, Javier - DAIS 2021
How Intuit uses Spark to Monitor In-Production Machine Learning Models at Large-Scale - SAIS 2020
How Not to Let Your Model and Data Drift Away Silently - Chengyin - MLOps 2021
- https://github.com/chengyin38/mlops_2021_drifting_away

Articles

Building a clinical data drift monitoring system with Azure DevOps, Azure Databricks, and MLflow - devblogs.microsoft - 2020-10-2

Azure ML

Data drift detection for datasets is currently in public preview - 2020-06-25
MLOps: Model management, deployment, and monitoring with Azure Machine Learning - 2020-03-17
Collect data for models in production - 2019-11-12
Detect data drift (preview) on datasets - 2020-06-25
Detect data drift (preview) on models deployed to AKS - 2019-11-04
Using Azure Machine Learning - Collect data from a scoring service - Microsoft lab
https://pypi.org/project/azureml-monitoring - Azure Machine Learning Data Collection API for Python
https://docs.microsoft.com/en-us/python/api/azureml-datadrift/azureml.datadrift?view=azure-ml-py
- https://docs.microsoft.com/en-us/python/api/azureml-datadrift/azureml.datadrift.datadriftdetector(class)?view=azure-ml-py

AWS Sagemaker

SageMaker Model Monitor

Amazon SageMaker Model Monitor – Fully Managed Automatic Monitoring For Your Machine Learning Models - blog - 2019-12-03
Introducing Amazon SageMaker Model Monitor – Maintain quality of ML models - 2019-12-03
Amazon SageMaker Model Monitor - sagemaker.readthedocs
Amazon SageMaker Model Monitor - doc
Monitoring a Model in Production - doc
Build, train & debug, and deploy & monitor with Amazon SageMaker - reinvent slides - 2019
https://github.com/awslabs/amazon-sagemaker-examples/blob/master/sagemaker_model_monitor/visualization/SageMaker-Model-Monitor-Visualize.ipynb
Blog
- Bring your own container to project model accuracy drift with Amazon SageMaker Model Monitor - 2021-07-26

Clarify

https://aws.amazon.com/sagemaker/clarify
- Identify and limit bias and explain predictions
- Clarify detects potential bias during data preparation, after model training, and in your deployed model by examining attributes you specify
- Detect bias in your data and model
  - Identify imbalances in data - integrated with Data Wrangler (feature engineering)
  - Check your trained model for bias
  - Monitor your model for bias
- Explain model behavior
  - Understand your model
  - Monitor your model for changes in behavior
  - Explain individual model predictions
- Use cases
  - Regulatory Compliance
  - Internal Reporting & Compliance
  - Customer Service
https://docs.aws.amazon.com/sagemaker/latest/dg/clarify-fairness-and-explainability.html
- Helps explain how these models make predictions using a feature attribution approach.
- Monitors inferences models make in production for bias or feature attribution drift.
https://sagemaker-examples.readthedocs.io/en/latest/sagemaker_processing/fairness_and_explainability/fairness_and_explainability.html
How Clarify helps machine learning developers detect unintended bias - 2020-12-09
https://github.com/aws/amazon-sagemaker-clarify
Articles
- https://techcrunch.com/2020/12/08/aws-announces-sagemaker-clarify-to-help-reduce-bias-in-machine-learning-models/ - 2020-12-08
- Introducing Amazon SageMaker Clarify — AWS re:Invent 2020 - julsimon - 2020-12-08

GCP

GCP Vertex AI

Introduction to Vertex Model Monitoring - Pre-GA Offerings Terms
Monitoring feature skew and drift

GCP AI Platform

Blogs

Event-triggered detection of data drift in ML workflows - Pre-GA Offerings Terms - 2021-03-15
Continuous model evaluation with BigQuery ML, Stored Procedures, and Cloud Scheduler - 2021-02-03
- Continuous evaluation—the process of ensuring a production machine learning model is still performing well on new data—is an essential part in any ML workflow.
- Performing continuous evaluation can help you catch model drift, a phenomenon that occurs when the data used to train your model no longer reflects the current environment.

Articles

https://dataintegration.info/monitor-models-for-training-serving-skew-with-vertex-ai - dataintegration - 2021-07-30
https://fuzzylabs.ai/blog/vertex-ai-the-hype/ - fuzzylabs.ai - 2021-06-30

Seldon

Drift Detection in Seldon Core

Drift Detection: An Introduction - 2021-07-14
- Monitoring and explainability of models in production - video - 2020-07-13

Monitoring and explainability of models in production - seminal PDF - 2020-07-13 -
- Production Machine Learning Monitoring: Outliers, Drift, Explainers & Statistical Performance - Saucedo - meetup - 2020-12-13 -

https://databricks.atlassian.net/wiki/spaces/~andre.mesarovic/pages/2045216594/Model+Serving+Resources#Seldon - Company basic
MLflow Integrations - MLflow Integrations Wiki page

Alibi Detect

Alibi Detect

github.com/SeldonIO/alibi-detect - Algorithms for outlier, adversarial and drift detection
- Jupyter examples
Alibi Detect - docs.seldon.io
- Seldon deployment of Alibi Outlier detector
seldonio/alibi-detect-server - Docker hub
Articles
- Simplifying Image Outlier Detection with Alibi Detect - Tolios - 2020-10-13

Articles

https://towardsdatascience.com/simplifing-image-outlier-detection-with-alibi-detect-6aea686bf7ba - Tollos - 2020-10-13

KFserving

Data Robot

Blogs
- All Data Drift is Not Created Equal - 2020-02-20
- Introducing MLOps Champion/Challenger Models - n.d.

Algorithmia

Algorithmia, Datadog Team on MLOps - Datanami 2020-11-05

ModelOp

Product
- Automated and Actionable Model Monitoring
  - Monitor all aspects of a model – operations, quality, risk and processes
  - Detect metrics and outcomes that exceed thresholds and controls
  - Identify incomplete steps and tasks in the model operations process
  - Initiate and track remediation steps until the problem is resolved
  - View the state and status of all models
ModelOp blog
- How to Monitor Your Machine Learning Models - 2020-11-18
- How to Ensure Better Model Performance Through ModelOps - 2020-07-16
Funding:
- Crunchbase - ModelOp
  - $6m - Seed - 2020-03-31

Whylabs

https://whylabs.ai

We take the pain out of model and data monitoring so that you spend less time firefighting, and more time building models.
WhyLabs enables them to operate with certainty by providing model monitoring, preventing costly model failures, and facilitating cross-functional collaboration.

https://whylabs.ai/blog

Data Logging: Sampling versus Profiling - 2020-10-28
whylogs: Embrace Data Logging Across Your ML Systems - 2020-09-22

Streamlining data monitoring with whylogs and MLflow - 2021-02-08
whylogs-examples/MLFlow Integration Example.ipynb - 2021-02-04

Sampling isn’t enough, profile your ML data instead - 2020-09-22
Introducing WhyLabs, a Leap Forward in AI Reliability - Visnjic - 2020-09-23

Crunchbase - whylabs
- $4m - Seed - 2020-09-23
Articles
- Amazon vets raise $4M from Madrona, Bezos Expeditions, others for AI2 spinout WhyLabs - geekwire - 2020-09-23

Alessya Visnjic - CEO - LinkedIn

Comet

https://www.comet.ml
Comet Model Production Monitoring (MPM) - focuses on models post production. The original product was more around how multiple offline experiments are modeled during training, while MPM is focused on these models once they hit production for the first time

Crunchbase - comet.ml
- pre-seed - Jul 17, 2017
- seed - Apr 5, 2018
- $4.5m - venture round - Apr 22, 2020
- $13m. - Round A - Apr 21, 2021
Articles
- MLOps startup Comet raises $13M to launch model monitoring products - venturebeat - 2021-04-08
- https://techcrunch.com/2021/04/08/comet-announces-13m-series-a-for-ml-model-building-tool/?guccounter=1&guce_referrer=aHR0cHM6Ly93d3cuZ29vZ2xlLmNvbS8&guce_referrer_sig=AQAAAH5v0-ufJ8VlwS3As-1tsA3RKd9S5_6SMN_7r8oL2yIuVSnSXlHiKfw7NDfNMgfZmSCV6Q6ofT8e02nCI5CeLbm_-is2Wdb5wkE_GZn_LTr91zmCsPRHan1RmFIP-S6P-eTDFf7ns408XlX9DFhJhM5Ijk4XN2_X6ct34mVoThNl - 2021-04-08

Gideon Mendels - CEO - founder
Nimrod Lahav - COO - founder

Evidently

https://github.com/evidentlyai/evidently
Evidently helps analyze machine learning models during development, validation, or production monitoring. The tool generates interactive reports from pandas DataFrame. Currently 6 reports are available.
- Data Drift - Detects changes in feature distribution.
- Numerical Target Drift - Detects changes in numerical target (see example below) and feature behavior.
- Categorical Target Drift - Detects changes in categorical target and feature behavior (see example below)
- Regression Model Performance - Analyzes the performance of a regression model and model errors (see example below).
- Classification Model Performance - Analyzes the performance and errors of a classification model. Works both for binary and multi-class models
- Probabilistic Classification Model Performance - Analyzes the performance of a probabilistic classification model, quality of model calibration, and model errors.
https://evidentlyai.com/blog/machine-learning-monitoring-data-and-concept-drift
Blog
- https://evidentlyai.com/blog/tutorial-3-historical-data-drift - 2021-07-29
  - example with Evidently, Plotly, Mlflow, and some Python code
  - evidently/historical_drift_visualization.ipynb - Example with MLflow

TorchDrift

https://torchdrift.org - TorchDrift is a data and concept drift library for PyTorch. It lets you monitor your PyTorch models to see if they operate within spec

Neptune

https://neptune.ai/blog/concept-drift-best-practices - 2021-04-29

Arthur.ai

https://www.arthur.ai - team
- Advanced monitoring and alerting, no matter where they’re deployed.
- Gain valuable insights into how your models are making decisions.
Blog
Documentation
Financials
- https://www.crunchbase.com/organization/arthur-ai
```
+------+--------+----------+
|amount|   round|      date|
+------+--------+----------+
| 18.3m|Total   |          |
| 15.0m|Series A|2020-12-09|
|  3.3m|Seed    |2019-12-11|
+------+--------+----------+
```
- https://github.com/art-ai
- Articles
  - https://techcrunch.com/2020/12/09/arthur-ai-snags-15m-series-a-to-grow-machine-learning-monitoring-tool/ - 2019-12-09
  - https://www.alleywatch.com/2020/12/arthur-ai-monitoring-platform-data-science-adam-wenchel/ - 2020-12
- Leaderhip
  - Adam Wenchel - CEO - co-cofounder - LinkedIn
  - John Dickerson - Chief scientists - co-cofounder - LinkedIn

Boxkite

Boxkite - capture feature and inference distributions used in model training, then compares them against realtime production distributions via Prometheus and Grafana.
https://boxkite.ml/en/latest/tutorials/kubeflow-mlflow/

Articles

Staying On Top of ML Model and Data Drift - datanami - 2020-07-16
anodot
- Learning the Learner: The Ultimate Way to Monitor Machine Learning
- https://github.com/anodot/MLWatcher - python agent that records a large variety of time-series metrics of your running ML classification algorithm
Hopswork
- https://github.com/logicalclocks/model-monitoring
Monitoring Machine Learning Models in Production - christophergs.com - 2020-03-14
Deployment Isn’t the Final Step – Monitoring Machine Learning Models in Production - imperva - 2019-11-25
ML-Powered Automatic Model Monitoring - medium - 2019-11-20
How to Monitor Machine Learning Models in Real-Time - Ted Dunning - kdnuggets - 2019-01
https://machinelearningmastery.com/gentle-introduction-concept-drift-machine-learning/ - 2017-12-15
https://monitorml.com - Real-Time Production Monitoring for Machine Learning
ModelOp Center - Monitoring capability - modelop.com - PDF

Other

https://en.wikipedia.org/wiki/Concept_drift#:~:text=In%20predictive%20analytics%20and%20machine,less%20accurate%20as%20time%20passes. - Wikipedia
Model Monitoring with R Markdown, pins, and RStudio Connect - Silge - 2021-04-08
https://github.com/scikit-multiflow/scikit-multiflow - Concept drift detection - amongst other things
amesar/mlflow-model-monitoring - MLflow model server monitoring and drift simple example (payload logging only)