Designing labor market recommender systems: the importance of job seeker preferences and competition

We examine the properties of a recommender system we developed at the Public Employment Service (PES) in France, prior to its implementation in the field. The algorithm uses past matches and a very large set of covariates to produce, for each job seeker, a ranking of the available offers and score each pair jobseeker-offer. Using a calibration step that takes advantage of the observation of application sequences, it gives a predicted "matching probability" for each pair.  After a theoretical discussion about the possible strategies to design a recommender system, we compare this new machine learning (ML) algorithm with another matching tool, mimicking the one currently used at the PES, based on a score measuring the "closeness" between the jobseeker's search criteria or preferences and the characteristics of the offer. We quantify the trade-off between the matching probability and the later "preference score" when switching from one system to the other. Next, we examine the issue of congestion.  We show that, on the one hand the ML algorithm based on past matches tends to increase congestion and on the other hand that this strongly reduces its performance. Finally, we show that the use of optimal transport to derive recommendations from the matching probability matrix significantly alleviates this problem. The main lesson at this stage is that an algorithm ignoring preferences and competition in the labor market would have very limited performances but that tweaking the algorithm to fit these dimensions substantially improves its properties, at least "in the lab".

Date

Speaker

Venue