Based on my 2 decades of experience helping enterprises adopt advanced analytics solutions, model drift is the largest reason for production model performance declines.
Businesses are able to move only a small share of their AI models to production. And then within 1-2 years, performance of most models deteriorate due to model drift. Businesses that can manage model drift will achieve multiple times more ROI from their models.
Explore why machine learning models degrade over time, or “drift”, and how you can detect and prevent it:
What is model drift?
Model drift, also called model decay, AI aging or temporal degradation, refers to the degradation of machine learning model performance over time. This means that the model suddenly or gradually starts to provide predictions with lower accuracy compared to its performance during the training period. According to recent research, 91% of ML models suffer from model drift.1
For example, suppose a hospital leverages a tool to predict if a patient might get a serious infection like sepsis. At first, the tool is helpful because it alerts doctors to treat patients early. But over time, the tool’s predictions start to fail. The model drift occurs because the tool doesn’t have this new data about what happens when doctors act on its warnings, so it starts making less accurate predictions.
| Drift Type | Example | Corrective Actions |
|---|---|---|
| Gradual concept Drift | Fraud detection evolves as criminals adapt new strategies over time, degrading the model's accuracy. | Perform feature engineering Retrain model |
| Sudden concept Drift | COVID-19 rapidly changed consumer spending habits, affecting demand forecasts trained on pre-pandemic data. | Revisit solution (new algorithm or approach) Evaluate solution relevance |
| Sudden concept drift: Reality | During a sudden surge in cases of a new respiratory illness, the model might fail to recognize the urgency due to lack of relevant training data, leading to inappropriate triage decisions. | Shift to human intervention Retrain model Reevaluate features |
| Recurring concept drift | Retail sales spike during holidays; a model that doesn't account for these seasonal changes predicts poorly. | Retrain model to account for seasonality Adjust model for recurring patterns |
| Data drift: Feature | A hospital model for predicting sepsis encounters new types of patient data that weren't in the training set. | Retrain model on new data Revisit feature engineering |
| Data drift: Label | A financial model misclassifies credit risk because the definition of "default" has changed. | Retrain on new data, tune |
| Data drift: Feedbac | A recommendation system overfits by prioritizing popular choices, leading to skewed suggestions over time. | Evaluate efficacy of solution Determine if a new solution is needed |
| Prediction Drift | Changes in regulations or market conditions impact financial models for credit risk or trading decisions. | Analyze impact on business |
What are the types of model drift?
There are two main types of model drift:
Concept drift
Concept drift happens when the relationship between input variables (or independent variables) and the target variable (or dependent variable) changes. This means that the definition of what we are trying to predict changes so that our algorithm provides inaccurate predictions.
Reasons for this can include latent features which are variables that have not been included in the model whose predictive value increases over time, rendering the model ineffective.
This change can be gradual, sudden, or recurring:
Figure 1. Concept drift detection framework.
Source: Knowledge-Based Systems 2
- Gradual concept drift: The change in fraudulent behavior is an example of gradual concept drift. As fraud detection methods become more sophisticated, fraudsters adapt to evade fraud detection systems by developing new strategies. An ML model trained on historical fraudulent transaction data would be unable to classify a new strategy as fraud. This means that the performance of the model would degrade because what is classified as fraud has changed over time.
- Sudden concept drift: Any sudden change in the environment can impact consumer behavior or global markets. For example, the 2021-2022 global chip shortage caused sudden concept drift, disrupting supply chain models that relied on stable conditions. These models failed to predict new supply constraints and demand surges, leading to operational challenges. Despite the shortage, global semiconductor sales grew by 25% in 2021, underscoring the need for companies to quickly adapt their models and strategies to manage the unforeseen scarcity and supply chain issues.3
- Recurring concept drift: This is also called seasonality. For instance, retail sales increase significantly during the Christmas season or on Black Friday. An ML model that does not take these known recurring trend changes into account would provide inaccurate predictions for these periods.
Data drift
Figure 2. Changing age distribution can cause data to drift.
Source: Evidently AI 4
Data drift occurs when the statistical properties of the input data change. For instance, as an online platform grows, the age distribution of its users may change over time. Since the usage habits of young and old people are not the same, a model trained on young people’s usage data would provide inaccurate predictions for old people’s behavior.
The impact of model drift on AI systems
- Inaccurate predictions or classifications: Model drift can lead to AI systems making less accurate predictions or classifications, as the model’s understanding of the data becomes outdated.
- Poor performance: As the model’s accuracy deteriorates, overall system performance can decline, which may affect the effectiveness of the AI system.
- Potentially harmful decisions: If the AI system is used in critical applications, such as healthcare or autonomous driving, inaccurate predictions can lead to harmful or unsafe decisions.
- System malfunction: In severe cases, model drift might cause the system to malfunction or behave unpredictably, which can have serious consequences.
- Financial losses: Inaccurate predictions and poor performance can result in financial losses, especially in business applications where decisions based on model outputs directly affect profitability.
- Physical harm: For AI systems used in safety-critical applications, such as industrial automation or transportation, model drift could potentially cause physical harm or safety issues.
How to deal with model drift?
Monitor model performance
In order to deal with model drift, data scientists should first be able to detect it before it causes major problems for end-users. Determining model performance metrics and continuously monitoring the performance of your model against them is therefore key to the long-term success of ML models.
Model monitoring solutions
There are specialized model monitoring tools and also MLOps platforms that provide model monitoring capabilities for drift detection. See MLOps tools for a selection of tools, and MLOps platforms for end-to-end platforms.
Check data quality
Some rapid performance changes can be due to problems in training data quality such as biases in data rather than concept or data drift. If that is the case, the problem would reveal itself early when you apply your model in a real-world use case. Check our article on bias in AI on different methods to fix it.
As an example, Google Health developed a deep learning model to detect a retina disease from patients’ eye scans.5 The model had 90% accuracy during its training phase but it failed to provide accurate results in real-life. This is because the model is trained with high-quality eye scans while real-world eye scans were lower in quality.
Ignoring a known seasonality is also a data quality issue. If your training data does not include recurring changes in the data, such as soaring retail sales during the Christmas season, this is a data quality issue that can be easily fixed.
Finally, while some mistakes in ground truth may not significantly impact model performance initially, they may lead to more rapid model drift. Therefore it is crucial to check the ground truth and ensure its quality.
Retrain the model
Figure 3. Periodically retraining the model can keep model quality high
Source: Databricks 6
If you detect a concept or data drift, you can apply model retraining with more recent data. Depending on the nature of the drift, there are different approaches:
- Use only recent data if old data has become outdated,
- Use all available data if the old data wouldn’t cause inaccurate model predictions,
- If the deployed model allows weighting, use all available data but assign higher weights to recent data so that the model pays less attention to old data.
Another option is online learning where the model continuously learns in real-time with the data feed. This will enable the model to keep itself up to date with evolving datasets.
Tune the model
If retraining the model doesn’t suffice, rebuilding the model can also help. This is because you have built your model with old training data in mind. Running multiple experiments with different features, hyperparameters, model architectures, etc. can help you update your model to keep in line with new data.
FAQ
Why is model drift unavoidable?
There is constant change in the business environment. These can be changing customer habits, economic pressures, or natural disasters, such as Covid-19. Therefore, the predictive accuracy of deployed machine learning models decrease over time.
For instance, consumer spending on recreational durable goods such as home fitness equipment increased by 18%, while spending on transportation services decreased by 23% in 2020.7 A demand forecasting model trained with pre-pandemic data would not predict these changes in consumer habits.
Further reading
Explore more on ML model improvement techniques such as:
External Links
- 1. Temporal quality degradation in AI models | Scientific Reports. Nature Publishing Group UK
- 2. Bayram, Firas & Ahmed, Bestoun & Kassler, Andreas. (2022). “From Concept Drift to Model Degradation: An Overview on Performance-Aware Drift Detectors.” Knowledge-Based Systems. 245. 10.1016/j.knosys.2022.108632. Retrieved at September 3, 2024.
- 3. Steward, Bucaille & Crossan (2021). “My kingdom for a chip: The semiconductor shortage extends into 2022.” Deloitte. Retrieved at September 3, 2024.
- 4. “What else can go wrong?” Evidently AI. Retrieved at September 3, 2024.
- 5. Heaven, D. (2020). “Google’s medical AI was super accurate in a lab. Real life was a different story.” MIT. Retrieved at September 3, 2024.
- 6. Heaven, D. (2020). Thomas & Mewled (2019). “Productionizing Machine Learning: From Deployment to Drift Detection” Databricks. Retrieved at September 3, 2024.
- 7. Barua (2021). “A spring in consumers’ steps: Americans prepare to get back to their spending ways.” Deloitte. Retrieved on September 3, 2024.
Comments
Your email address will not be published. All fields are required.