Economic recommendation based on pareto efficient resource allocation

Abstract

A fundamentally important role of the Web economy is Online Resource Allocation (ORA) from producers to consumers, such as product allocation in E-commerce, job allocation in freelancing platforms, and driver resource allocation in P2P riding services. Since users have the freedom to choose, such allocations are not provided in a forced manner, but usually in forms of personalized recommendation, where users have the right to refuse. Current recommendation approaches mostly provide allocations to match the preference of each individual user, instead of treating the Web application as a whole economic system where users therein are mutually correlated on the allocations. This lack of global view leads to Pareto inefficiency, i.e., we can actually improve the recommendations by bettering some users while not hurting the others, and it means that the system did not achieve its best possible allocation. This problem is especially severe when the total amount of each resource is limited, so that its allocation to one (set of) user means that other users are left out. In this paper, we propose Pareto Efficient Economic Recommendation (PEER) - that the system provides the best possible (i.e., Pareto optimal) recommendations, where no user can gain further benefits without hurting the others. To this end, we propose a Multi-Objective Optimization (MOO) framework to maximize the surplus of each user simultaneously, and provide recommendations based on the resulting Pareto optima. To benefit the many existing recommendation algorithms, we further propose a Pareto Improvement Process (PIP) to turn their recommendations into Pareto efficient ones. Experiments on real-world datasets verify that PIP improves existing algorithms on recommendation performance and consumer surplus, while the direct PEER approach gains the best performance on both aspects.