Industrial recommender systems usually consist of the retrieval stage and the ranking stage, to handle the billion-scale of users and items. The retrieval stage retrieves candidate items relevant to user interests for recommendations and has attracted much attention. Frequently, users show hierarchical multi-interests reflected in a heavy user of a certain NBA team Golden State Warriors in Sports, who is also a light user of almost the whole Animation. Both Sports and Animation are at the same level. However, most existing methods implicitly learn this hierarchical difference, making more fine-grained interest information to be averaged and limiting detailed understanding of the user's different needs in heavy interests and other light interests. Therefore, we propose a novel two-stage approach to explicitly modeling hierarchical multi-interest for recommendation in this work. In the first hierarchical multi-interest mining stage, the hierarchical clustering and transformer-based model adaptively generate circles or sub-circles that users are interested in. In the second stage, the partition of retrieval space allows the EBR models to only deal with items within each circle and accurately capture user's refined interests. Experimental results show that the proposed approach achieves state-of-the-art performance. Our framework has also successfully deployed at Lofter (one of the largest derivative content communities with 10 million monthly active users) for over four months.

翻译：工业推荐系统通常由检索阶段和排序阶段组成，以应对十亿级别的用户和物品规模。检索阶段负责检索与用户兴趣相关的候选物品进行推荐，因此受到广泛关注。用户通常表现出分层多兴趣特征——例如某位重度关注体育领域NBA球队金州勇士的用户，同时也是动画领域的轻度用户。此处体育与动画属于同一层级。然而，现有方法大多隐式学习这种层级差异，导致更细粒度的兴趣信息被平均化，难以深入理解用户在重度兴趣与其它轻度兴趣上的差异化需求。为此，本文提出一种新颖的两阶段方法，显式建模推荐中的分层多兴趣。在第一阶段的分层多兴趣挖掘中，基于层次聚类和Transformer的模型可自适应生成用户感兴趣的兴趣圈层或子圈层。第二阶段通过检索空间划分使EBR模型仅需处理各圈层内物品，从而精准捕获用户的精细化兴趣。实验结果表明，所提方法达到了当前最优性能。该框架已在Lofter（月活跃用户超千万的头部衍生内容社区）成功部署逾四个月。