Abstract
This paper questions the design of job recommender systems (RS). We argue that a direct application of sophisticated Machine Learning (ML) algorithms to build recom- mendations, such as identifying offers most likely to lead to a job from the prediction of successful matches, does not always lead to an improvement for job seekers (JS). This is because the objectives of these recommendations do not align with the ones of the JS and they are usually generated independently of each other, without taking into account the competition. Using a theoretical model of two-sided market with a step of applications, we show that the ML tools from which the recommendations are di- rectly derived can be more usefully mobilized to identify quantities that JS might have difficulties to access. When JS have biased expectations, our main result is that a RS combining the chances of being hired and the utility of the jobs, coming closer to the expected utility, would dominate the others in identifying offers that JS would not have thought of themselves and making search more productive. Our empirical analysis con- firms this quantitatively matters, using the RS designed inside of a long-term project we are conducting with the French Public Employment Service, which leverages rich and detailed data on applicants, firms, and past job searches. We also discuss how RS can avoid increasing congestion in using a collective objective rather that an individ- ual one to generate the recommendations, using optimal transport to make it tractable.
The paper is joint with Guillaume Bied, Philippe Caillou, Christophe Gaillac, Elia Perennes, Michele Sebag
Invited by: Giulio Zanella
Local Organizer: Giovanni Angelini