EN FR
EN FR


Section: Application Domains

Computational Social Sciences: Toward AI Fairness

Several TAU projects are related to computational social and economic sciences. This activity is at the core of the French DataIA Institut de Convergence, (head Nozha Boujemaa), gathering 19 partners in the Paris-Saclay area to explore the scientific and ethical impacts of data science and artificial intelligence on the academic, industrial and societal sectors.

Many projects in the domain are related to Causal Modelling (see Section 7.1.1). Some are internal to our team; others involve collaborations with external partners, with a transfer dimension. Others are closely related to some Software platform and are desribed in the corresponding Sections (io.datascience, Section 6.1 and Catolabe, Section 6.3).

  • AmiQap (Philippe Caillou, Isabelle Guyon, Michèle Sebag, Paola Tubaro, started 2015). The multivariate analysis of state questionaire data relative to the quality of life at work, in relation with the socio-economical indicators of firms, aims at investigating the relationship between quality of life and economic performances (conditionally to the activity sector), in collaboration with the RITM (U. Paris-Sud), SES (IMTelecom) and La Fabrique de l'Industrie, on data gathered by the Ministry of Labour (DARES). AmiQap is a motivating application for the Causal Modelling studies (PhD Divyan Kalainathan; post-doc Olivier Goudet; coll. David Lopez-Paz, Facebook AI Research).

  • Collaborative Hiring (Philippe Caillou, Michèle Sebag, started 2014). Thomas Schmitt's PhD, started in 2014, aims at matching job offers and resumes viewed as a collaborative filtering problem. An alternative approach based on Deep Networks has been developped by François Gonard within his IRT SystemX PhD. The study has been conducted in cooperation with the Web hiring agency Qapa and the non-for-profit organization Bernard Gregory.

  • U. Paris-Saclay Nutriperso IRS (Philippe Caillou, Flora Jay, Michèle Sebag, Paola Tubaro) aims to uncover the relationships between health, diets and socio-demographic features. The ultimate goal is to provide personalized acceptable recommendations toward healthier eating practices. A milestone is to uncover the causal relationships between diet and health (coll. INRA, INSERM, CEA).

  • RESTO (Paola Tubaro, Philippe Caillou). A study of transformations brought about by digital platforms and their effects on the restaurants sector, using a mix of methods that includes both agent-based simulations and machine learning, and fieldwork.

  • Sharing Networks (Paola Tubaro, started 2016). Mapping the "collaborative economy" of internet platforms through social network data and analysis.

  • OPLa - DiPLab (Paola Tubaro). Two related projects investigating the economy of micro-work platforms in France, and how they integrate with the AI industry ecosystem.

Scientific challenges are related to the FAT (Fairness, Accountability and Transparency) criteria: Metric learning, where the distance/topology to be learned must reflect prior knowledge (e.g. ontologies); Interpretation of clusters built from heterogeneous textual and quantitative data, using the learnt metric/distance; Integration of the human-in-the-loop ("dire d'experts"); Assessment of the models w.r.t. their causality (as opposed to their predictive accuracy) in order to support further interventions.