Course contentsShow
← Machine Learning and Deep Learning
Lesson 438 of 3,53810. Feature EngineeringFree lesson

Handling Outliers: Removal, Capping, and Transformation

Strategies for dealing with outliers including winsorization, trimming, and robust transformations like log scaling.

Handling Outliers: Removal, Capping, and Transformation

You'll learn three practical strategies for managing outliers once you've detected them in your data.

The Core Idea

After identifying outliers (using methods from lesson 436-437), you need to decide what to do with them. Simply detecting outliers isn't enough—you must handle them appropriately. The three main strategies are:

1. Removal (Trimming)

Delete rows containing outliers entirely. This works when outliers represent data errors or you have abundant data. However, removing too much data can hurt model performance.

Analogy: Like throwing out spoiled fruit from a basket—simple but wasteful if you discard too much.

2. Capping (Winsorization)

Replace extreme values with a threshold (usually the 5th and 95th percentiles, or 1st and 99th). Values beyond these limits are "capped" at the boundary.

Example: If income above 500Kiscappedat500K is capped at 500K, a billionaire's income becomes $500K in your dataset.

This preserves the number of samples while limiting extreme influence.

3. Transformation

Apply mathematical functions to compress the scale and reduce outlier impact. Common transformations include:

  • Log transformation: log(x) reduces right-skewed data
  • Square root: sqrt(x) for moderate skewness
  • Box-Cox transformation: automatically finds the best power transformation

Analogy: Like adjusting camera exposure to see both bright and dark areas—transformations bring outliers closer to the main distribution without discarding information.

Which Strategy to Choose?

  • Use removal sparingly when outliers are clearly errors
  • Use capping when you need to preserve all samples
  • Use transformation when outliers are legitimate but need taming, especially before algorithms sensitive to scale

Key Takeaway: The three outlier-handling strategies—removal, capping, and transformation—each preserve different aspects of your data; choose based on whether outliers represent errors, legitimate extremes, or skewed distributions.