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

Bagging classifier starts by separating the training data into smaller sets randomly created. It is important to note that the same event can be present in different sets. This final group of sets is called a bootstrapped dataset. Decision trees are formed for each one of the subsets. The model is then a combination of those trees. The prediction will be the mode of the predictions. The third method is called random forest and is similar to bagging but with one variation. During the process of creating the decision trees, instead of creating a rule based on one of the total number of variables, two variables will be selected randomly and the decision will be made between those variables. This process is done for each branch. As for the simple decision tree classifier the maximum depth and minimum number of samples of a leaf will be optimized but for those methods the number of trees/ estimators will also be tuned. It is important to note that overfitting of the estimators is not a problem as the combination of them annulates those variations. The last method is histogram-based gradient boosting . This method creates decision trees sequentially boosting the importance of points depending on previous error. The training process starts with the creation of a tree with one leaf, comparable to a dummy classifier, a model that will always predict the most common class. The residual is calculated for each element and a new tree will be created with the objective of predicting the residuals. The new predictions are now a sum of the dummy classifier with the predicted residuals multiplied by a learning rate. This process repeats many times in order to improve the quality of the result. As this method improves from previous mistakes this is considered machine learning. In order to optimize gradient boosting the same parameters will be tuned as for bagging and random forest, but the impact is different. Gradient boosting generally thrives with underfitting decision trees with each one correcting the previous ones. Figure 20 : Representation of simple decision tree 28