An Unbiased Data Collection and Content Exploitation/Exploration Strategy for Personalization.

摘要: One of missions for personalization systems and recommender is to show content items according users' personal interests. In order achieve such goal, these are learning user interests over time trying present tailoring profiles. Recommending preferences has been investigated extensively in the past few years, mainly thanks popularity Netflix competition. a real setting, users may be attracted by subset those interact with them, only leaving partial feedbacks system learn next cycle, which leads significant biases into hence results situation where engagement metrics cannot improved time. The problem not just one component system. data collected from usually used many different tasks, including ranking functions, building profiles constructing classifiers. Once biased, all downstream use cases would impacted as well. Therefore, it beneficial gather unbiased through interactions. Traditionally, collection done showing uniformly sampling pool. However, this simple scheme feasible risks takes long feedbacks. paper, we introduce user-friendly framework, utilizing methods developed exploitation exploration literature. We discuss how framework normal multi-armed bandit problems why method needed. layout novel Thompson Bernoulli ranked-list effectively balance experiences collection. proposed validated bucket test strong comparing old algorithms