Abstract: Monotonic classification is an ordinal classification problem in which the monotonic constraint exists between features and class. There have been some methods which can deal with the monotonic classification problem on the nominal datasets well. But for the monotonic classification problems on the numeric datasets, the classification accuracies and running efficiencies of the existing methods are limited. In this paper, a monotonic classification method based on decision forest (MCDF) is proposed. A sampling strategy is designed to generate decision trees, which can make the sampled training data subsets having a consistent distribution with the original training dataset, and the influence of non-monotonic noise data is avoided by the sample weights. It can effectively improve the running efficiency while maintaining the high classification performance. In addition, this strategy can also determine the number of trees in decision forest automatically. A solution for the classification conflicts of different trees is also provided when the decision forest determines the class of a sample. The proposed method can deal with not only the nominal data, but also the numeric data. The experimental results on artificial, UCI and real datasets demonstrate that the proposed method can improve the monotonic classification performance and running efficiency, and reduce the length of classification rules and solve the monotonic classification problem on large datasets.

Key words: monotonic classification, decision tree, monotonic consistency, decision forest, ensemble learning