Distribution Shift

What is it?

Distribution shift happens when a Machine Learning model training data differs from the one encountered in a production environment. This causes the model to perform poorly when deployed, even though it can achieve good results at the training phase.

For example, a time series forecasting problem needs to continuously train the model with recent data. If not, then the model can yield worse results as time passes, because of the effect of distribution shift on production data.

This issue is present, to a small degree, in almost every Machine Learning application, because of the difficulty to perfectly replicate production conditions at training time.

---

Types of distribution shift

There are three main types of distribution shift, that may happen simultaneously depending on the context.

---

Covariate shift

Covariate shift occurs when $p(x)$ changes between training and production data, but $p(y|x)$ does not. In other words, the features may change its distribution, but the relationship between the features and target stay the same regardless of training or production data.

For example, a facial recognition model trained with European-only faces may perform poorly when deployed in other continents, even though the relationship $p(y|x)$ stay the same.

---

Concept shift

Concept shift occurs when $p(y|x)$ changes between training and production data, but $p(x)$ does not. It is quite the opposite of covariate shift. Now, the relationship between features and target change, but the data distribution stays the same.

This is a harder shift to detect, because most times will be related to the business side of the problem, not the model itself.

For example, an accurate travel recommendation model might have performed poorly when the pandemic hit, because of an external factor closing borders around the world. This is not inherent to the model, but the context of deployment itself.

---

Label shift

Label shift occurs only in situations where $y\to x$ ($y$ is believed to cause $x$), occurring when $p(y)$ changes between training and production data but $p(y|x)$ does not.

In other words, the target/ labels might change, but the relationship between the features and target stay the same.

---

Detecting distribution shifts

To detect distribution shifts in a deployed model, one need to have a solid MLOps workload. Most commonly, distribution shifts happens gradually, so one must continually monitor and evaluate the performance of a deployed model.

In general, the most common ways to detect distribution shifts is to monitor model performance and data.

• Monitoring model performance

Evaluate performance metrics from time to time, like accuracy and precision, or any other statistical or evaluation metric. It these change over time, it may be due to distribution shift.

• Monitor data

It’s also possible to evaluate data shift by comparing statistical properties of training and production data. It also may prove valuable to store historical properties of the data.

---

Dealing with distribution shifts

In general, fixing data issues — collecting a better training set or applying Data-Centric AI methods — and re-training the model can address most issues.

To address concept shift, one must have a solid business understanding of the problem which the model is deployed. External factors might be in play and only evaluating the model might not prove efficient.

---

References

• Class Imbalance, Outliers and Distribution Shift — MIT — Introduction to Data-Centric AI — Course.

• Dataset Shift in Machine Learning — Book.