Feature Engineering and Feature Scaling

Three people, two columns: age and salary. Ask which two are more similar, and the answer of a distance-based algorithm changes when the salary is written in thousands instead of single units. Same people, different answer. This video explains why that happens and how feature scaling repairs it, after a short first half on what a feature is and how to build a better one. You will learn: What a feature is, and why a well-built feature often helps more than a cleverer model (length x width = area) Why scale: a computed distance example with age and salary, and why gradient descent struggles in a narrow valley Standardisation by hand on five numbers: subtract the mean, divide by the standard deviation (population version) Min-max normalisation on the same five numbers What one outlier does to both methods Why the scaler is fitted on the training data only (data leakage) Which models need scaling, and why decision trees do not Chapters 0:00 Who is more similar to Asha? 0:43 What a feature is 1:14 Building a better feature 1:55 Why scale: distance 3:15 Why scale: gradient descent 3:43 Standardisation, by hand 4:50 Min-max normalisation 5:50 What an outlier does 6:34 Fit on the training data only 7:18 Which models need scaling 7:55 Recap + try it yourself Try it yourself: standardise the numbers 1, 3, 4, 5 and 7 with the population standard deviation, then apply min-max scaling to the same numbers. Animations made with Manim Community Edition. The narration is a synthetic voice. #MachineLearning #FeatureEngineering #AIEngineer More from Harban AI Labs: https://harban-ai-labs.vercel.app/