Crowd Avoidance and Diversity in Socio-Economic Systems and Recommendation

Recommender systems recommend objects regardless of potential adverse effects of their overcrowding. We address this shortcoming by introducing crowd-avoiding recommendation where each object can be shared by only a limited number of users or where object utility diminishes with the number of users sharing it. We use real data to show that contrary to expectations, the introduction of these constraints enhances recommendation accuracy and diversity even in systems where overcrowding is not detrimental. The observed accuracy improvements are explained in terms of removing potential bias of the recommendation method. We finally propose a way to model artificial socio-economic systems with crowd avoidance and obtain first analytical results. Introduction. – Recommender systems are a powerful tool which nowadays helps most online retailers to make effective offers to their consumers. They use past user preferences to recommend new objects that the users might like. Research of recommendation grows rapidly and tackles issues like recommendation algorithms [1, 2], recommendation in social systems [3], and the use of recommendation in e-commerce [4]. While there are situations where an arbitrary number of users can be recommended the same object, in other situations this is not the case. For example, one cannot recommend the same restaurant to many people as it has limited space and service capabilities. Similarly, it is not advantageous to use data on industrial production in countries [5–7] and recommend the same new product to many countries as it could lead to undue competition and a poor ultimate outcome. We propose crowd-avoiding recommendation which addresses this issue by imposing a strict occupancy constraint on individual objects or by assuming that object utility decays with the number of users sharing it. Our approach is linked to physics where particles occupy the energetically most favorable states but are either allowed in single occupation (fermions) or obey no restrictions (bosons). Although for particles there are no other options, we are here interested also in situations lying between these two extremes where an object can be recommended to an intermediate number of users. There is also a close connection with the combinatorial assignment problem where agents (users) can perform certain tasks (objects) with a certain cost and one searches for a bijective agent-task matching that minimizes the sum of the corresponding costs [8]. We use here some of the algorithms originally developed for the assignment problem. Note that crowd avoidance is a general concept which can be used also in situations where resources can be shared by an arbitrary number of parties and user satisfaction does not decay with the number of other users sharing the resource. Given a set of user score (or cost) values, one can always apply an occupancy constraint or penalty and see how this impacts the assignment of objects to users. It is of particular importance to note that this new assignment is bound to be more diverse than the original one where no additional constraints were present. For example, if an individual object scores top for many users, it cannot be assigned to all of them if the occupancy constraint is sufficiently strong. Other objects then have to replace it and the composition of the assigned objects becomes more diverse. In this way, the crowd avoidance concept can help address one of the long standing challenges of information filtering: the lack of diversity [9] and its potential adverse impact on network topologies [10].