Beyond Two-Tower: Attribute Guided Representation Learning for Candidate Retrieval

Candidate retrieval is a key part of the modern search engines whose goal is to find candidate items that are semantically related to the query from a large item pool. The core difference against the later ranking stage is the requirement of low latency. Hence, two-tower structure with two parallel yet independent encoder for both query and item is prevalent in many systems. In these efforts, the semantic information of a query and a candidate item is fed into the corresponding encoder and then use their representations for retrieval. With the popularity of pre-trained semantic models, the state-of-the-art for semantic retrieval tasks has achieved the significant performance gain. However, the capacity of learning relevance signals is still limited by the isolation between the query and the item. The interaction-based modeling between the query and the item has been widely validated to be useful for the ranking stage, where more computation cost is affordable. Here, we are quite initerested in an demanding question: how to exploiting query-item interaction-based learning to enhance candidate retrieval and still maintain the low computation cost. Note that an item usually contain various heteorgeneous attributes which could help us understand the item characteristics more precisely. To this end, we propose a novel attribute guided representation learning framework (named AGREE) to enhance the candidate retrieval by exploiting query-attribute relevance. The key idea is to couple the query and item representation learning together during the training phase, but also enable easy decoupling for efficient inference. Specifically, we introduce an attribute fusion layer in the item side to identify most relevant item features for item representation. On the query side, an attribute-aware learning process is introduced to better infer the search intent also from these attributes. After model training, we then decouple the attribute information away from the query encoder, which guarantees the low latency for the inference phase. Extensive experiments over two real-world large-scale datasets demonstrate the superiority of the proposed AGREE against several state-of-the-art technical alternatives. Further online A/B test from AliPay search servise also show that AGREE achieves substantial performance gain over four business metrics. Currently, the proposed AGREE has been deployed online in AliPay for serving major traffic.