Covariate Shift
Covariate shift refers to the situation where the distribution of $X_{\text {test }}$ changes from $\mathcal{P}$ to $\mathcal{P}_{\text {test }}$, but the relationship between $X_{\text {test }}$ and $Y_{\text {test }}$, i.e. the distribution of $Y_{\text {test }} \mid X_{\text {test }}$, stays fixed.
The key insight here is that while the marginal distribution P(X) shifts, the fundamental relationships between features and outcomes P(Y|X) remain stable across domains.
Examples
Examples where the input distribution changes but the conditional relationship Y|X remains the same:
Medical Diagnosis
• X: Patient features (age, symptoms, test results, demographics)
• Y: Diagnosis outcome (heart attack, no heart attack)
• Training: Model trained on patients from urban hospitals (younger, more diverse population)
• Test: Applied to rural hospitals (older, less diverse population)
• X changes: Age distribution shifts older, different ethnic composition
• P(Y|X) stays same: A 65-year-old with chest pain and elevated troponin has the same probability of heart attack regardless of hospital location
Spam Detection
• X: Email features (words, sender, links, formatting, urgency indicators)
• Y: Email classification (spam, not spam)
• Training: Emails from 2020 mentioning "COVID vaccines" and "work from home"
• Test: Emails from 2025 mentioning "AI tools" and "remote collaboration"
• X changes: Vocabulary and topics shift with current events
• P(Y|X) stays same: Urgent language + suspicious links + requests for personal info still indicate spam with the same probability
House Price Prediction
• X: Property features (size, location, bedrooms, school district, neighborhood type)
• Y: House price
• Training: Suburban houses in California
• Test: Urban apartments in New York
• X changes: Property types, neighborhood characteristics, size distributions
• P(Y|X) stays same: A 2-bedroom, 1000 sq ft property with good schools nearby has a predictable price relationship regardless of location