Affiliations: [a] Faculty of Engineering of The University of Porto, Porto, Portugal | [b] Artificial Intelligence and Computer Science Laboratory, Porto, Portugal
Correspondence:
[*]
Corresponding author: Xavier Fontes, Faculty of Engineering of The University of Porto, Rua Dr. Roberto Frias s/n 4200-465 Porto, Portugal. E-mail: [email protected].
Note: [1] The first author was supported by the Calouste Gulbenkian Foundation, under a New Talents in Artificial Intelligence Program Grant.
Abstract: This paper introduces Evolutionary Directed Graph Ensembles (EDGE). EDGE combines ideas from social dynamics and trust in human beings with graph theory. We use pre-trained prediction models as nodes in a directed acyclic graph where the connections between nodes have associated weight matrices to simulate the trust each node has in its predecessors. EDGE uses a genetic algorithm approach to evolve a population of these directed acyclic graphs, in an ensemble-type hybridization process. The pre-trained models are stored in a pool of models named Reservoir. The Reservoir can be populated with models from different families, such as Decision Trees, Ensemble Methods or Neural Networks. To test EDGE, four datasets were used: a dataset of a parking lot occupancy taken from a university student parking lot; a dataset about appliances energy use in a low energy building; a dataset of Anuran calls; and the MNIST dataset. Results show that we can achieve good accuracy measures of around 98% for the MNIST dataset, 99% for the Anuran data, 86% on the Appliances Energy and about 88% on the parking lot dataset. EDGE proves to be robust against bad performing nodes, presenting average accuracy results of up to 49% above the worst performing node in the ensemble. It also never shows results below the best performing node, and in some cases even improves the results with respect to the best node by up to 4%.
