This paper proposes a new tree-based ensemble method for supervised classification and regression problems, called Extra-Trees (ET). The method randomizes both attribute and cut-point selection during tree node splitting. In the extreme case, it builds totally randomized trees whose structures are independent of the output values of the learning sample. The strength of randomization can be tuned by a parameter. The paper evaluates the robustness of the default parameter choice and provides insights on how to adjust it. The main strength of the algorithm is computational efficiency. A bias/variance analysis of the ET algorithm is provided, along with a geometrical and kernel characterization of the models induced. The ET algorithm builds an ensemble of unpruned decision or regression trees using a top-down procedure. It differs from other tree-based ensemble methods by fully randomizing cut-point and attribute selection during splitting. The algorithm uses the full learning sample rather than bootstrap replicas to grow trees. The ET algorithm has parameters K (number of attributes randomly selected at each node), n_min (minimum sample size for splitting a node), and M (number of trees in the ensemble). These parameters can be adjusted to problem specifics. The paper presents an empirical evaluation of the ET algorithm on 24 datasets, comparing it with other tree-based methods such as CART, Tree Bagging, Random Subspace, and Random Forests. The results show that ET performs as well as or better than these methods in terms of accuracy and computational efficiency. The ET algorithm is particularly effective in classification problems, where it outperforms other methods. The algorithm's performance is also robust to high noise conditions. The paper also analyzes the effect of the parameters K, n_min, and M on the performance of the ET algorithm. It shows that the default values of these parameters are generally robust and effective. The analysis of the bias/variance characteristics of the ET algorithm shows that it has a better trade-off between bias and variance than other tree-based methods. The ET algorithm is also shown to be effective in regression problems, although it may not perform as well as other methods in some cases. The paper concludes that the ET algorithm is a promising method for supervised learning, with good accuracy and computational efficiency. It is particularly effective in classification problems and is robust to high noise conditions. The algorithm's performance is also influenced by the choice of parameters, which can be adjusted to problem specifics. The paper provides insights into the behavior of the ET algorithm and its effectiveness in different scenarios.This paper proposes a new tree-based ensemble method for supervised classification and regression problems, called Extra-Trees (ET). The method randomizes both attribute and cut-point selection during tree node splitting. In the extreme case, it builds totally randomized trees whose structures are independent of the output values of the learning sample. The strength of randomization can be tuned by a parameter. The paper evaluates the robustness of the default parameter choice and provides insights on how to adjust it. The main strength of the algorithm is computational efficiency. A bias/variance analysis of the ET algorithm is provided, along with a geometrical and kernel characterization of the models induced. The ET algorithm builds an ensemble of unpruned decision or regression trees using a top-down procedure. It differs from other tree-based ensemble methods by fully randomizing cut-point and attribute selection during splitting. The algorithm uses the full learning sample rather than bootstrap replicas to grow trees. The ET algorithm has parameters K (number of attributes randomly selected at each node), n_min (minimum sample size for splitting a node), and M (number of trees in the ensemble). These parameters can be adjusted to problem specifics. The paper presents an empirical evaluation of the ET algorithm on 24 datasets, comparing it with other tree-based methods such as CART, Tree Bagging, Random Subspace, and Random Forests. The results show that ET performs as well as or better than these methods in terms of accuracy and computational efficiency. The ET algorithm is particularly effective in classification problems, where it outperforms other methods. The algorithm's performance is also robust to high noise conditions. The paper also analyzes the effect of the parameters K, n_min, and M on the performance of the ET algorithm. It shows that the default values of these parameters are generally robust and effective. The analysis of the bias/variance characteristics of the ET algorithm shows that it has a better trade-off between bias and variance than other tree-based methods. The ET algorithm is also shown to be effective in regression problems, although it may not perform as well as other methods in some cases. The paper concludes that the ET algorithm is a promising method for supervised learning, with good accuracy and computational efficiency. It is particularly effective in classification problems and is robust to high noise conditions. The algorithm's performance is also influenced by the choice of parameters, which can be adjusted to problem specifics. The paper provides insights into the behavior of the ET algorithm and its effectiveness in different scenarios.