A review and comparison of strategies for handling missing values in separate-and-conquer rule learning

In this paper, we review possible strategies for handling missing values in separate-and-conquer rule learning algorithms, and compare them experimentally on a large number of datasets. In particular through a careful study with data with controlled levels of missing values we get additional insights on the strategies’ different biases w.r.t. attributes with missing values. Somewhat surprisingly, a strategy that implements a strong bias against the use of attributes with missing values, exhibits the best average performance on 24 datasets from the UCI repository.