RINA - Démonstrateur d’utilisation de l’intelligence artificielle pour la gestion opérationnelle des risques naturels gravitaires géologiques | FEREC 2020

Figure 29: heat mapping of the performance of decision tree model as the hyperparameters varie for models with classes with different weights. Performance of model measured by (a) (b) β=1 balanced accuracy Similar to the heat maps on Figure 27, Figure 29 shows a great performance for trees with minimum depth and high amount of samples per leaf. On the other hand, there seems to be an optimum model for a tree with a depth of 12 and min_samples_leaf equal to 7. Either option, max_depth = 3 with min_samples_leaf = 9 or max_depth = 12 with min_samples_leaf = 7, could be the best model. Both were calculated and the second one overfitted considerably. The balanced accuracy for the training set was 0.74 and the was equal to 0.56 while the test set β=1 was considerably worse (balanced accuracy = 0.63 and = 0.4). The model with lower depth β=1 had a similar performance and the results for the training and testing sets were much closer. Actual 0 1 Predicted 0 867 99 1 77 47 Figure 30: Confusion matrix representing the result of predictions based on the decision tree classifier method for the train set. 38