Whether collected all along the lifetime of a product or acquired through interactions with a user, temporal data provide the principal source of information in industry and services for decision and planning.
To name but a few, predictive maintenance, user modeling, network management and supply forecasting are all motivations for developing predictive tools able to deal with multiple sensors, unlabeled data, rare events, noisy environments and time-to-detection constraints. For that purpose, the following lines of research will be deepened:
The growing availability of vast quantities of data about users and their environment, notably conveyed by the Web and the Internet of Things, open the door to their exploitation for designing new services to the user/customer. If recommendation, classification, regression and ranking can naturally underpin those services, difficulties essentially arise from the lack of supervision, the heterogeneity of the data sources and their large volume.
Indeed, mostly unlabeled or partially labeled, user data require to be handled in new settings such as few-shot learning and weakly supervised learning that go beyond the classical framework of unsupervised or supervised learning.
Additionally, heterogeneity in data modalities raises the difficulty to represent them into a common space, a hurdle that can be overcome by metric and representation learning. Eventually more complex predictive models (opinion summarization, automatic captioning…) can be developed. Regarding these goals, our group aims at developing the following lines of research:
Diagnosis and decision in critical environments call for advanced machine learning tools with new guarantees such as correctness, traceability and interpretability of decision, outlier awareness, robustness to adversarial inputs and ability to abstain in order to give the hand to the human expert at the right moment. In some contexts, additional properties such as privacy and fairness are also of great interest.
Researches on this area are still in their infancy and the development of methods with these new guarantees will take place through important changes in the Machine Learning paradigm. The following research directions will be explored:
Whether it be installed on a cloud or embedded onto a device, new AI systems are expected to interact with a changing environment (a face recognition system in an airport, a self-driving car, a smart network) and to update accordingly their decision function or their database if needed. A necessary step towards autonomous systems consists in thinking learning as a continuous process that not only exploits information from environment but also keeps on exploring it appropriately, and eventually monitor itself.
Beyond the paradigm of reinforcement and online learning, especially devoted for that purpose, better interaction with the environment also take place through a relevant definition of the real value function for the target user (quality of experience, for instance). Working directions will mainly concern the following topics: