What problem does this paper attempt to address?

The problem that this paper attempts to solve is: how to use domain expert knowledge to enhance the detection and mitigation of concept drift in machine learning (ML) models during actual operations. ### Problem Background With the wide application of machine learning models in modern software systems, changes in data distribution (such as changes caused by external events and data integrity issues) may lead to a decline in model performance. This phenomenon is known as **data drift**, which includes: - **Feature Drift**: changes in the input data distribution \(P(X)\). - **Feature Drift**: changes in the input data distribution \(P(X)\). - **Concept Drift**: changes in the conditional probability distribution \(P(Y|X)\) of the target variable given the input. Concept drift has a significant impact on model performance, especially in sensitive areas such as credit card fraud detection, and may lead to discriminatory behavior. Therefore, timely detection and mitigation of concept drift are crucial. ### Existing Challenges 1. **Concept Drift Detection in Unlabeled Data**: In the real world, labeled data is often delayed or missing, making detection based on metrics such as accuracy difficult. 2. **Understanding the Characteristics of Detected Concept Drift**: Even if concept drift can be detected, it is still challenging to understand its characteristics such as severity, frequency, duration, and transition speed. 3. **Limitations of Automated Tools**: Existing automated concept drift detection and mitigation tools are not commonly used and rely more on human intervention and shift - based monitoring. ### Solution To solve the above problems, this paper proposes the **Expert - Driven Monitoring** method, which specifically includes: 1. **Scenario Specification**: - Through expert knowledge acquisition and retrospective analysis, collect events that may cause concept drift and standardize them. - Use the Bayesian model to quantify the experts' subjective beliefs about changes in feature distribution and estimate parameters such as mean and standard deviation. 2. **Scenario Identification**: - After detecting feature drift at runtime, use the Bayesian model to compare and infer the occurrence of scenarios. - Calculate the posterior probability \(P(M|D)\) of each scenario model and compare it with the reference model to obtain the Bayesian factor to evaluate the likelihood of the scenario. 3. **Response Mechanism**: - Select the most appropriate response strategy based on the Bayesian factor to assist ML engineers in making decisions or automatically trigger mitigation measures. By integrating the knowledge of domain experts, this method aims to improve the detection accuracy and response efficiency of concept drift, thereby ensuring the stability and reliability of machine learning models in practical applications.

Expert-Driven Monitoring of Operational ML Models

A Model-Driven Engineering Approach for Monitoring Machine Learning Models

A novel lifelong machine learning-based method to eliminate calibration drift in clinical prediction models.

How to Sustainably Monitor ML-Enabled Systems? Accuracy and Energy Efficiency Tradeoffs in Concept Drift Detection

MLDemon: Deployment Monitoring for Machine Learning Systems

Time to Retrain? Detecting Concept Drifts in Machine Learning Systems

Learning Run-time Safety Monitors for Machine Learning Components

Designing monitoring strategies for deployed machine learning algorithms: navigating performativity through a causal lens

A monitoring framework for deployed machine learning models with supply chain examples

Monitoring machine learning (ML)-based risk prediction algorithms in the presence of confounding medical interventions

Considerations for Quality Control Monitoring of Machine Learning Models in Clinical Practice

Explanatory Model Monitoring to Understand the Effects of Feature Shifts on Performance

Enhancing Model Adaptability Using Concept Drift Detection for Short-Term Load Forecast

Leveraging Expert Consistency to Improve Algorithmic Decision Support

AlerTiger: Deep Learning for AI Model Health Monitoring at LinkedIn

Automating concept-drift detection by self-evaluating predictive model degradation

Safety Monitoring of Machine Learning Perception Functions: a Survey

Into the Unknown: Active Monitoring of Neural Networks

Monitoring Machine Learning Forecasts for Platform Data Streams

Incorporating Experts' Judgment into Machine Learning Models

Runtime Monitoring of Human-centric Requirements in Machine Learning Components: A Model-driven Engineering Approach