Contrastive Learning for Debiased Candidate Generation at Scale

Deep candidate generation (DCG), which narrows down the enormous corpus to a few hundred candidate items via representation learning, is integral to industrial recommender systems. Standard approaches adopt maximum likelihood estimation (MLE) and rely on sampling to ensure scalability, which reduces DCG to a task similar to language modeling. However, live recommender systems face severe unfairness of exposure with a corpus several orders of magnitude larger than that of natural language, which implies that (1) MLE will preserve and even exacerbate the exposure bias in the long run, as it aims to faithfully fit the history records, and (2) suboptimal sampling and inadequate use of item features can lead to inferior representations for the items that are unfairly ignored. In this paper, we introduce CLRec, a Contrastive Learning paradigm successfully deployed in a real-world massive recommender system, for alleviating exposure unfairness in DCG. We theoretically prove that a popular choice of contrastive loss is equivalent to reducing the exposure bias via inverse propensity scoring, which complements previous understanding of contrastive learning. We further implement a good sampling distribution and reuse most of the computation when encoding rich features for both positive and negative items, by employing a fix-sized queue to store items (and reuse the computed representations) from previous batches, where the queue serves as the negative sampler. In contrast, previous state-of-art DCG approaches do not use rich features for the negative items due to high computational costs, since the number of negative samples needs to be large. Extensive offline analyses and online A/B tests in Mobile Taobao demonstrate substantial improvement, including a dramatic reduction in the Matthew effect.