Vadore – Using Data to Improve Job Search
Vadore (“Valorisation des données pour la recherche d’emploi”) is a long-term collaboration between CREST, LISN and France Travail.
It has received funding from DataIA, Dares and J-PAL.
The project’s goal is to design, test and scale up job-matching recommender systems for the labor market. In the long run, we aim to build tools that:
recommend vacancies to jobseekers, and
recommend suitable candidate profiles to employers,
in a way that is effective, scalable, and fair.
We started by designing recommender systems targeted at jobseekers.
A key question is: what kind of jobs should be recommended?
Jobs that are very close to the search criteria (“a job I like”), or jobs with high hiring chances (“a job I can get”)? Our work shows that aligning the nature of recommendations with jobseekers’ expectations matters a lot.
To answer this, we ran several beta tests with jobseekers. These tests followed a learning cycle: we formulate hypotheses, test them in the field, analyze the results, then update the algorithm and test again. This process is essential to design tools that truly fit jobseekers’ needs.
Main scientific outputs:
Toward Job Recommendation for All
IJCAI 2023, AI and Social Good Track (PDF link)
A Job I Like or a Job I Can Get: Designing Job Recommender Systems Using Field Experiments (working paper, submitted 2025)
– Working paper available from Christophe Gaillac’s research page (PDF link):
Registered experiments (AEA RCT Registry):
Designing Job Recommendation Algorithms for Job Seekers: First Beta Test
Bied et al., 2022 (Link)
Designing Job Recommendation Algorithms for Job Seekers: Second Beta Test
Bied et al., 2025 (Link)
The second step, currently underway, is to test the impact of the algorithm at scale.
We run a large randomized evaluation in which jobseekers receive weekly recommendations. The design is organized by micro-markets, which also allows us to study congestion (what happens if many people are pushed toward the same vacancies):
in some micro-markets, about 90% of jobseekers receive Vadore recommendations every week;
in others, only 25% are exposed;
in pure control markets, no one receives Vadore recommendations;
in additional markets, 25% receive recommendations from the existing France Travail recommender, allowing for direct comparison.
Registered experiment (AEA RCT Registry):
Testing Recommender Systems with Jobseekers in France
Nathan et al., 2025 (Link)
The ongoing randomized experiments as received funding from DARES.
A key risk with recommender systems is congestion: if the same few attractive vacancies are recommended to everyone, competition explodes and overall performance can collapse.
We therefore develop recommender algorithms that explicitly control congestion, in particular through optimal transport:
instead of optimizing recommendations independently for each jobseeker,
we use a global objective that spreads applications more evenly across vacancies.
This helps both to avoid over-crowding on a few offers and to give visibility to “orphan” vacancies that receive no applications.
Main scientific outputs:
Congestion-Avoiding Job Recommendation with Optimal Transport
FEAST workshop, ECML-PKDD 2021 (Link)
Designing Labor Market Recommender Systems: the Importance of Job Seeker Preferences and Competition(Link)
Randomized experiments are planned as part of the DARES funding to test the impact of such congestion avoiding recommender systems.
We also study fairness, focusing on gender differences in recommendations:
Do men and women receive recommendations that differ in terms of hiring chances, contract type, wages, etc.?
Can we reduce unfair gaps without destroying the overall performance of the system?
We document gender gaps, propose interpretation tools, and test post-processing and adversarial de-biasing methods to align average performance across groups.
Main scientific output:
Fairness in Job Recommendations: Estimating, Explaining, and Reducing Gender Gaps
AEQUITAS Workshop at ECAI 2023 (link)
Randomized experiments are planned as part of the DARES funding to test the impact of such fairness-oriented adjustments in practice.
Finally, Vadore also looks at recommendations on the firm side: recommending candidate profiles to employers.
Previous work with France Travail has shown that firms are very sensitive to the quality of recruitment services, especially when the public employment service helps pre-screen candidates and reduces the flood of unsuitable applications.
This suggests a high potential for tools that:
send small sets of highly relevant candidates to firms,
reduce firms’ recruitment frictions,
and possibly increase overall hiring.
Related evaluation:
Are Active Labor Market Policies Directed at Firms Effective? Evidence from a Randomized Evaluation with Local Employment Agencies
Y. Algan, B. Crépon, D. Glover – Rev. of Economic Studies, conditonnaly accepted (link)
This line of work extends the Vadore tools to recommend jobseeker profiles to firms, and is supported by a dedicated J-PAL grant.