EDGE: Evolutionary Directed Graph Ensembles

Classification tasks are being tackled in a plethora of scientific fields, such as astronomy, finance, healthcare, human mobility, and pharmacology, to name a few. Classification is defined as a supervised learning approach that uses labeled data to assign instances to classes. A common approach to tackle these tasks are ensemble methods. These are methods that employ a set of models, instead of just one and combine the predictions of every model to obtain the prediction of the whole. Common obstacles in ensemble learning are the choice of base models to use and how best to aggregate the predictions of each individual to produce the ensemble’s prediction. It is also expected to mitigate the weaknesses of its members while pooling their strengths together. It is in this context that Evolutionary Directed Graph Ensembles (EDGE) thrives. EDGE is a machine learning tool based on social dynamics and modeling of trust in human beings using graph theory. Evolutionary Algorithms are used to evolve ensembles of models that are arranged in a directed acyclic graph structure. The connections in the graph map the trust of each node in its predecessors. The novelty in such an approach stems from the fusion of ensemble learning with graphs and evolutionary algorithms. A limitation of EDGE is that it focuses only on changing the topology of the graph ensembles, with the authors of hypothesizing about using the learned graphs for other tasks. with gains as substantial as 30 percentage points. The bootstrap was shown to be effective in improving the prediction power, with the exploitation of previous runs improved the results on 19 out of 21 datasets. The contributions can be summarized as a novel way to evolve graph ensembles, by also evolving the weights between nodes of the graphs, coupled with the idea of bootstrapping any dataset using previous runs from other datasets. The analysis of dataset choice for the bootstrapping lead to the proposal of a similarity metric between datasets that can be used to facilitate the choice for bootstrapping, without exhaustive or random search in the available datasets. uma métrica de semelhança que pode ser utilizada em vez de uma pesquisa exaustiva nos conjuntos de dados disponíveis.