Method for Hyperparameter Tuning of Image Classification with PyCaret

Abstract

A method for hyperparameter tuning of image classification with PyCaret is proposed. The application example compares 14 classification methods and confirms that Extra Trees Classifier has the best performance among them, AUC=0.978, Recall=0.879, Precision=0.969, F1=0.912, Time=0.609 bottom. The Extra Trees Classifier produces a large number of decision trees, similar to the random forest algorithm, but with random sampling of each tree and no permutation. This creates a dataset for each tree containing unique samples, and from the ensemble set of features a certain number of features are also randomly selected for each tree. The most important and unique property of the Extra Trees Classifier is that the feature split values are chosen randomly. Instead of using Gini or entropy to split the data to compute locally optimal values, the algorithm randomly selects split values. This makes the tree diverse and uncorrelated. i.e. the diversity of each tree. Therefore, it is considered that the classification performance is better than other classification methods. Parameter tuning of Extra Trees Classifier was performed, and training performance, test performance, ROC curve, accuracy rate characteristics, etc. were evaluated.