DMC

Supervised classification is becoming increasingly promising in several real world applications such as precision medicine, condition monitoring or anomaly detection. From attributes that can be numeric, categorical, images or any other kind of signals describing each instance, the objective is to predict the true class of samples in order to support the experts of the application domain in their diagnosis. However, three important issues often occur in such real-world applications: (i) the presence of imbalanced classes which generally leads most of classifiers to be inaccurate for detecting the smallest classes, (ii) the presence of prior probability shifts which damages the classification error risk of most fitted classifiers when the class proportions evolve, (iii) the presence of covariate shifts which damages the classification error risk of every fitted classifier when the distribution of the attributes change between the training set and test samples. According to the literature of statistical decision theory, the minimax classifier is theoretically designed to be the most robust face to imbalanced datasets and prior probability shifts since it is a Bayesian classifier that minimizes the maximum of the class-conditional risks. In our previous research, we developed a new approach for computing a multi-class minimax classifier suitable to process both numeric and categorical attributes, that can be coupled with any kind of Deep Neural Network for processing images. The first objective of our present proposal project is now to address the issue of covariate shifts in order to make the minimax classifier more robust when dealing with uncertainty in the feature distributions. Secondly, we also aim to extend these research to multi-label classification tasks in which these three issues often occur too in safety critical applications, which is especially difficult when dealing with mixed features.