Zero Rule

Zero Rule (ZeroR) is a benchmark procedure for classification algorithms.

Zero Rule algorithms do not take the features of a data point into account, but only follow the distribution of labels observed during training. It is used often used to provide a baseline performance by predicting the most frequent class in the dataset.

For example, if 65% of the data items belong to a particular class A, ZeroR will predict that class for all data items, achieving an accuracy of 65%. Alternatively, it can assign labels randomly based on the distribution of labels in the training data. In the above example, every datapoint has a 65% change of being labeled as A. As randomness is involved, the observed overall accuracy can vary but approaches 65%.

ZeroR is simple and effective as a benchmark: if an algorithm performs worse than ZeroR, it is not useful for the given domain.