Section: Research Program

Causal modelling

Data science is viewed as an information processing cycle: i) exploiting data and prior knowledge to build models; ii) using models to support optimal decisions in view of desirable ends; and iii) acquiring more data in order to refine the models and/or the desirable ends. Inasmuch data science supports prescriptive recommendations, it requires building causal models: these hold in front of interventions on the application domain $-$ as opposed to predictive models. Causal modelling, acknowledged a priority at the international level (DARPA 2015), opens principled and sound ways to deal with the unbounded expectations / irrational exuberance about Big Data. Furthermore, causality offers an operational framework to better handle transfer learning, semi-supervised learning and missing data.

Mainstream approaches to causality involve restrictive assumptions (no confounders; no causal cycles) with severe scalability limitations [83]. The international challenges proposed by I. Guyon in the last 3 years (See book in preparation) opened brand new research directions, based on learned causation models [78], [75]. The validation of causal graphs still is an open problem in the general case (multiple hypothesis testing issues, heterogeneous variables, temporal dimensions).

TAU is one of the first teams worldwide with expertise in this domain, collaborating with Max Planck Institute (B. Schölkopf), FORTH (I. Tsamardinos) and Facebook Research (D. Lopez Paz). Among the applications calling for causal models are Energy Management (RTE use-cases include failures of equipment and catastrophic cascades of failures; Inria post-doc work of Berna Batu, see Section 4.1) and computational social sciences (with impact on strategic societal issues, see Section 4.2).