AI Hints
Statistics & Math for ML · Lesson 23 of 138

Min-Max Scaling

Squeeze into 0 to 1.

⏱ 2 min read▶ Video at 25:33🎓 Free
Watch this lesson in the Statistics & Math for ML courseWatch on YouTube ↗Full Statistics & Math for ML course

Sometimes you want every feature squeezed between zero and one. For example, pixel brightness in images, or inputs to some neural networks. That's min max scaling.

Min-Max Scaling
From the video: Min-Max Scaling
  • Squeeze into 0 to 1
  • Pixels, neural networks
  • Min-max scaling

The formula

Subtract the minimum, and divide by the range. The smallest value becomes zero, and the largest becomes one. Everything else lands in between, in proportion. Run it. Twenty, thirty five, and fifty become zero, a half, and one.

minmax.pypython
import numpy as np

ages = np.array([20, 35, 50, 26, 44])

scaled = (ages - ages.min()) / (ages.max() - ages.min())
print(scaled.round(2))
Output
[0.  0.5 1.  0.2 0.8]

Which scaler?

Standardizing is the usual default. It's less affected by a single extreme value. Min max scaling gives a fixed range, which is nice for images and some neural networks, but one outlier can squash everything else into a tiny range. Tree models, like random forests, don't need scaling at all.

StandardScaler

  • mean 0, std 1the usual defaulthandles outliers better

MinMaxScaler

  • range 0 to 1images, some networksoutliers squash the rest

Summary

Min max scaling maps the smallest value to zero and the largest to one. Standardizing is usually the safer default. Next, a quiz on spread and scaling.

Key takeaways
  • (x − min) / (max − min)
  • 0 to 1
  • Outliers can squash it
Next up: Spread Quiz — test yourself