UML-NET

Modern information systems become more and more complex, and so do the data they have to handle. This is the case, for example, of protein functions prediction, information retrieval from big data bases of documents, recommending items or actions from few preferences, diagnosing and monitoring the state of complex systems such as planes, energy networks, … This means that accurate data and models are respectively harder to obtain and harder to learn. Hence, available data are often noisy and incomplete (e.g., pairwise instead of full preferences). Dealing with such data while preserving the computational efficiency of the learning methods often means a drop in the accuracy of model predictions. While some loss of accuracy for a gain of efficiency is affordable in some applications such as recommender systems or web crawlers, such a loss is a tremendous drawback in other applications such as medical diagnosis, risk analysis or autonomous vehicles. For these latter applications, recent and very promising trends of research consist in fully acknowledging this uncertainty in the learning process, possibly sacrificing some efficiency for more reliable results, for instance by allowing models to make partial and cautious but more accurate predictions, or by treating data uncertainty in a conservative way. The goal of this project is to foster collaborations between French laboratories and European research groups following this trend from different perspectives, to exchange views, establish a common understanding of this emerging field and develop a European network focusing on this specific topic. The aim of this project if to lay down the first corner stones of this long-term goal, notably through the orgnization of events (workshops, special sessions or journal special issues) and the exchange of researchers and students between the different involved groups and the Heudiasyc Laboratory. The network will focus on two main topics. The first topic addressed by the network the one learning from uncertain data, either in the input features or the observed (structured) output. In particular, it will explore the recent trend recommending to learn sets of models from the uncertain data, thereby producing partial but more reliable prediction, and will compare it to the more usual requirement of learning one model giving a precise prediction. The second explored topic will concern cautiousness in models and in decision-making, more specifically when cautiousness is understood as the fact of producing sets of models or sets of predictions when uncertainty is too important. Two problems related to this we expect to discuss are how this cautiousness should be exploited in learning methods, and how to handle this cautiousness in a computationally efficient way.