Fast online learning to recommend a diverse set from big data

Abstract

Building a recommendation system to withstand the rapid change in items’ relevance to users is a challenge requiring continual optimization. In a Big Data scenario, it becomes a harder problem, in which users get substantially diverse in their tastes. We propose an algorithm that is based on the UBC1 bandit algorithm to cover a large variety of users. To enhance UCB1, we designed a new rewarding scheme to encourage the bandits to choose items that satisfy a large number of users. Our approach takes account of the correlation among the items preferred by different types of users, in effect, increasing the coverage of the recommendation set efficiently. Our method performs better than existing techniques such as Ranked Bandits [8] and Independent Bandits [6] in terms of satisfying diverse types of users.