A dynamic multiple classifier system using graph neural network for high dimensional overlapped data

Dynamic selection techniques select a subset of the classifiers from a pool according to their perceived competence in labeling each given query instance in particular. To do so, most techniques rely on the locality assumption for the selection task, meaning that similar instances should share a set of adequate classifiers, so their competencies are usually estimated over a local region surrounding the query. However, as the local distribution is crucial to these techniques, a poor region definition due to the presence of high dimensionality and class overlap can have a negative impact on their performance, thus limiting their application. Thus, we propose in this work a dynamic selection technique to better deal with sparse and overlapped data in which the instance-instance and the classifier-classifier relationships are leveraged to learn the dynamic classifier combination rule. The proposed technique uses a multi-label graph neural network as a meta-learner, so both the data modeled as a graph, without directly defining the local region, and the classifiers' inter-dependencies modeled in the meta-labels are used to learn an embedded space where the dynamic selection task is more straightforward. Experimental results over 35 high dimensional datasets show that the proposed method significantly outperforms the static selection baseline and most evaluated dynamic selection techniques when using a diverse ensemble. Moreover, the proposed technique surpassed the contending state-of-the-art techniques over the problems with the highest excess of incompetent classifiers in overlap regions, further suggesting its suitability to deal with challenging local distributions. Code available at: github.com/marianaasouza/gnn_des.