Four labeled training cards are bootstrap-sampled with replacement to build separate tree models, showing how each sample leaves certain cards out of training. Those left-out cards become out-of-bag data, scored only by trees that never saw them. The animation derives the exclusion probability (1-1/n)^n, computing the exact value 81/256 for n=4 and contrasting it with the large-n limit of about 0.368. It then extends bagging to random forests via random feature selection at splits, and cautions that OOB scores are not a fully untouched test after repeated tuning. Useful for students learning ensemble methods and model evaluation.
Narrated · 16:9 · Preview before teaching · automatic layout checks do not establish subject accuracy
# Bagging trains on overlapping samples and evaluates what was left out Four training cards labeled 1, 2, 3, and 4 are sampled with replacement. The first bootstrap sample is [1,1,3,4], leaving sample 2 out of that model's training. A second sample [2,2,2,4] leaves samples 1 and 3 out. Train separate tree icons from these samples and show how each observation receives out-of-bag predictions only from trees that did not train on it. Derive the exclusion probability (1-1/n)^n: for n=4 it is 81/256, while the large-n limit is approximately .368. Keep that limit distinct from the exact small example. Combine predictions through voting, then introduce random forests by adding random candidate-feature selection at tree splits. Explain that out-of-bag evaluation is useful but is not an untouched final test after repeated tuning. Use black backgrounds, readable sample identities, and clearly separated training branches. Audience: 一年级人工智能专业学生