Joint BCAM-UPV/EHU Data Science and Artificial Intelligence seminar: Efficient Learning of Minimax Risk Classifiers in High Dimensions

Link to the session here

Abstract
High-dimensional data is common in multiple areas, such as health care and genomics, where the number of features can be hundreds of thousands. In such scenarios, the large number of features often lead to inefficient learning. In addition, the conventional performance assessment based on cross-validation increases the computational cost and can be unreliable when the number of samples is significantly lower than the number of features. Recently, methods based on constraint generation have enabled efficient learning of L1-regularized support vector machines (SVMs). This talk will present such methods to obtain an efficient learning algorithm for the recently proposed minimax risk classifiers (MRCs). The presented iterative algorithm obtains a sequence of MRCs with decreasing worst-case error probabilities while learning. Therefore, the algorithm can address the trade-off between training time and the classifier performance that can be suitable for the scenarios discussed above. In addition, the algorithm also provides a greedy feature selection as a side benefit.