KuaiRand: An Unbiased Sequential Recommendation Dataset with Randomly Exposed Videos

Abstract

Recommender systems deployed in real-world applications can have inherent\nexposure bias, which leads to the biased logged data plaguing the researchers.\nA fundamental way to address this thorny problem is to collect users'\ninteractions on randomly expose items, i.e., the missing-at-random data. A few\nworks have asked certain users to rate or select randomly recommended items,\ne.g., Yahoo!, Coat, and OpenBandit. However, these datasets are either too\nsmall in size or lack key information, such as unique user ID or the features\nof users/items. In this work, we present KuaiRand, an unbiased sequential\nrecommendation dataset containing millions of intervened interactions on\nrandomly exposed videos, collected from the video-sharing mobile App, Kuaishou.\nDifferent from existing datasets, KuaiRand records 12 kinds of user feedback\nsignals (e.g., click, like, and view time) on randomly exposed videos inserted\nin the recommendation feeds in two weeks. To facilitate model learning, we\nfurther collect rich features of users and items as well as users' behavior\nhistory. By releasing this dataset, we enable the research of advanced\ndebiasing large-scale recommendation scenarios for the first time. Also, with\nits distinctive features, KuaiRand can support various other research\ndirections such as interactive recommendation, long sequential behavior\nmodeling, and multi-task learning. The dataset and its news will be available\nat https://kuairand.com.\n