Most multi-label learning (MLL) techniques perform classification by analyzing only the physical features of the data, which means they are unable to consider high-level features, such as structural and topological ones. Consequently, they have trouble to detect the semantic meaning of the data (e.g., formation pattern). To handle this problem, a high-level framework has been recently proposed to the MLL task, in which the high-level features are extracted using the analysis of complex network measures. In this paper, we extend that work by evaluating different combinations of four complex networks measures, namely clustering coefficient, assortativity, average degree and average path length. Experiments conducted over seven real-world data sets showed that the low-level techniques often can have their predictive performance improved after being combined with high-level ones, and also demonstrated that there is no a unique measure that provides the best results, i.e., different problems may ask for different network properties in order to have their high-level patterns efficiently detected.
Poor Reconstruction of Complex Network Measures From M/EEG