This animation contrasts feature selection, which keeps a subset of original columns, with feature extraction, which projects columns into new combined axes. It visualizes a 20-feature search space of over a million subsets, compares filter, wrapper, and embedded selection workflows, and uses the XOR problem to show how individual features can look useless while their combination carries full class information. A final pipeline diagram shows correct train-only fitting of scaling and selection to avoid test-set leakage, useful for first-year AI students.
Narrated · 16:9 · Preview before teaching · automatic layout checks do not establish subject accuracy
# Selecting features means choosing information, not creating new axes A data table shows useful, redundant, constant, and noisy feature columns. Feature selection keeps some original columns, while a separate projection animation combines columns into new coordinates to illustrate feature extraction. Display the search space of 20 features as 2^20 possible subsets, including the empty set. Compare filter, wrapper, and embedded selection through three concise workflows. Then show the four equally likely XOR input combinations: either binary feature alone carries no information about the class, but the pair determines it. This reveals why ranking individual features can miss interactions. Draw a training-and-validation pipeline in which scaling and feature selection are fitted only inside the training portion of each split. Keep the test set outside the selection process. Use black backgrounds, cyan feature highlights, orange rejected columns, and short English labels. Useful for understanding relevance, redundancy, interaction, and leakage. Audience: 一年级人工智能专业学生