Completely Lazy Classiers: Bayesian Neighborhoods
msra
摘要
Local classiers are sometimes called lazy learners because they do not process the train- ing data until presented with a test sample. However, such methods are generally not completely lazy, because the neighborhood size k (or other locality parameter) is usually chosen by cross-validation on the training set, which can require signicant preprocess- ing and risks overtting. We propose a simple Bayesian alternative to cross-validation of the neighborhood size that requires no pre-processing. We motivate this Bayesian neigh- borhoods technique by showing it extends the standard Bayes decision rule to minimize expected misclassication costs by treating the neighborhood size as a random variable. Additionally, the proposed estimated posterior minimizes the expected Bregman loss. We analyze the eect
更多查看译文
关键词
cross-validation,bayesian qda,local learning,bayesian estimation,lazy learning
AI 理解论文
溯源树
样例
生成溯源树,研究论文发展脉络