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

Figure 19: representation of data set to illustrate nearest neighbors model The parameter that was tuned for each one of the different types of models is the number of neighbours that would be taken into consideration and weighting strategies. 4.1.2. Decision Tree This model consists in a series of binary logical decisions in order to choose which category it will predict. Each decision is based simply on one of the features. One example is, if the rain of the previous day was superior to 30mm. This rule can lead to other branches or to the final decision. Four different types of decision tree models were explored. The first and simplest is a simple decision tree classifier. During the training stage the model will create rules that maximize performance. Adjusting the value of the hyperparameters can help to avoid the model overfitting, fitting too perfectly to the train set and not only finding general tendencies, or underfitting, not being able to find the general tendencies of the model. Those parameters are the maximum depth of the trees, the minimum number of samples in each leaf and the weights of each class. The maximum depth of the model stops it from growing excessively and overfitting. The minimum number of samples of a leaf stops branches extending with few samples once more avoiding the model overfitting. The class weights is a form of giving more importance to detecting events. As for the third variation of the nearest neighbors it is interesting to give more importance to predicting the maximum number of events than predicting an event that did not happen. If both classes had the same weight and a leaf had 21 days with no event and 10 with events it would predict there would not be an event. If the weight of the class with rock falls was 3 times larger it would predict that there would be an event. The other decision tree classifiers, bagging , random forest and histogram-based gradient boosting, are produced by combining different trees. The idea is to create a more flexible model. 27