The maximal information coefficient is a measure that was proposed in 2011 and can detect non-linear relationships in experiments using artificial data. However, its effectiveness on real-world data has not been sufficiently demonstrated. In this study, various benchmark data sets from different fields were gathered to evaluate the effectiveness of the maximal information coefficient in real-world classification tasks. Distance-based discriminant analysis and support vector machine were adopted as classifiers. Accuracies and computational costs were employed to evaluate the results, Compared to the baselines including Euclidean distance, the Pearson correlation coefficient, cosine similarity and Spearman's rank correlation coefficient, the classification accuracy of the maximal information coefficient failed to show superiority and its computational costs were significantly higher than the other measures.
展开▼