Automatic Classifier Selection Based on Classification Complexity

Choosing a proper classifier for one specific data set is important in practical application. Automatic classifier selection (CS) aims to recommend the most suitable classifiers to a new data set based on the similarity with the historical data sets. The key step of CS is the extraction of data set feature. This paper proposes a novel data set feature that characterizes the classification complexity of problems, which has a close connection with the performance of classifiers. We highlight two contributions of our work: firstly, our feature can be computed in a low time complexity; secondly, we theoretically show that our feature has connection with generalization errors of some classifiers. Empirical results indicate that our feature is more effective and efficient than the existing data set features.