Leave-one-out cross-validation is a special case of cross-validation in which one record is left out of the build data when building a model. The number of models built equals the number of records (omitting a different build record for each model), which makes this procedure computationally expensive. With Naive Bayes models, however, the approach can be modified such that all build records are used for building a single model. Then, the model is repeatedly modified to quickly remove the effects of one build record, incrementally "unbuilding" the model for that record, as though that record had been omitted when building the model in the first place. The accuracy of the prediction for each build record can then be assessed against the model that would have been built from all the build records except that one, without having had to actually build a separate model for each build record.
