RECTO : REcommandation diminuant la Congestion par Transport Optimal
Résumé
The matching setting, a particular case of recommendation problem, focuses on applications where a so-called item can be attributed to at most one user, with the job market and the matrimonial market as chief examples. In such cases, recommending the items preferred by each user might contribute to a congestion issue as users aiming at the same item cannot be all satisfied. While some state of art approaches proceed by repairing the recommendation
policy to account for the congestion issue, other approaches take inspiration from the optimal transport (OT) framework, and aim to map the user population onto the item population in order to minimize some global transportation cost. In OT-based recommendation state-of-art approaches,
the collaborative matrix (reporting the user-item matches) is interpreted as if it were the result of an OT plan ; the underlying transport cost is inferred and used e.g. to propose new congestion-avoiding recommendation policies. In this paper, another OT-based recommendation strategy is defined, noting that the collaborative matrix reflecting the individual decisions can hardly be considered as the result of a (centralized) OT plan. Accordingly, the proposed algorithm first learns a recommendation score from the data and then defines a cost matrix, with the transportation cost decreasing depending on the matching relevance. The optimal transport plan is thereafter used for the recommendation. The experimental validation of the approach presents com-
parative results on benchmark data on the matrimonial market, and proprietary real-world data on the job market, illustrating the trade-off between the recall and the congestion indicators. The discussion focuses on the integration of individual and collective criteria.
La recommandation de biens rivaux (sur le marché du travail ou des rencontres) fait face au danger majeur de la congestion. Pour concevoir un système de recommandation évitant la congestion, une approche possible repose sur le transport optimal, cherchant un appariement global entre l'ensemble des utilisateurs et l'ensemble des items selon un coût de transport à définir. L'originalité de l'approche RECTO (REcommandation diminuant la Congestion par
Transport Optimal) est de construire le coût de transport en fonction du score de recommandation, et de définir la politique de recommandation en fonction du plan de transport entre utilisateurs et items. Une validation expérimentale comparative, menée sur une base de données publique
relative au marché matrimonial et une base de données propriétaire relative au marché du travail, illustre le compromis entre précision et congestion. La discussion porte sur l’intégration de critères individuels et collectifs.
| Origine | Accord explicite pour ce dépôt |
|---|---|
| licence |
