Multivariate Outlier Detection
What you'll learn: How to spot outliers that hide in multiple dimensions using relationships between variables.
The Limitation of Single-Variable Methods
So far, you've learned to detect outliers by examining one variable at a time—using z-scores, IQR, or MAD. But what if a data point looks normal in every individual dimension yet is still unusual when you consider variables together?
Analogy: Imagine measuring height and weight separately. A person might have a typical height and typical weight individually, but the combination (very tall but very light, for example) could be unusual.
Enter Mahalanobis Distance
Mahalanobis distance measures how far a point is from the center of a multivariate distribution, accounting for correlations between variables. Unlike simple distance measures, it stretches and shrinks with the shape of your data cloud.
Think of it as asking: "How weird is this combination of values, given how these variables normally relate to each other?"
Why Multivariate Relationships Matter
Variables rarely exist in isolation. Height and weight correlate. Temperature and ice cream sales correlate. When you ignore these relationships, you miss outliers that are unusual in context.
Multivariate outlier detection examines the entire pattern: not just "is this value extreme?" but "is this combination of values extreme given what we know about how they usually vary together?"
When to Use It
Use multivariate outlier detection when:
- You have multiple related numerical variables
- Relationships between variables matter for your analysis
- Single-variable methods miss contextually unusual observations
Key Takeaway: Multivariate outlier detection using Mahalanobis distance finds data points that are unusual in the context of multiple correlated variables together, catching anomalies that single-variable methods miss.