Combining individual models together with the purpose of improving the predictive power of the model.
Card 1 of 259. Answer: Combining individual models together with the purpose of improving the predictive power of the model.
About · Quizzes · History of AI · To suggest new questions or report an error send me a dm.
Ensemble learning combines the predictions of several models, called base learners, into one prediction that is usually more accurate and stable than any single member. The members can be one algorithm trained on different data samples or different algorithms, merged by voting, averaging or learned weights.
It works because errors that are not perfectly correlated partly cancel. Take three binary classifiers, each 70% accurate, that err independently. A majority vote is right when at least two are: 0.7³ + 3 · 0.7² · 0.3 = 0.784. If all three made the same mistakes, the vote would stay at 70%: diversity is what an ensemble buys.
Interviewers expect the main families. Bagging trains models in parallel on bootstrap samples and mainly cuts variance; boosting trains them in sequence on earlier errors and mainly cuts bias; stacking feeds out-of-fold predictions to a meta-model. The costs: slower inference and harder explanations. Netflix never deployed the full blend that won its 2009 prize, judging the gain not worth the engineering effort.
