Simple but Efficient: A Multi-Scenario Nearline Retrieval Framework for Recommendation on Taobao

TL;DR

This work tackles the scalability-latency tension in the matching stage of recommender systems by introducing a Multi-Scenario Nearline Retrieval (MNR) framework that reuses fine-grained ranking signals from multiple Taobao scenarios in near real-time. It aggregates cross-scenario ranking results via Flink, maintains per-scenario histories with FIFO queues, and applies a streaming scoring rule to produce a compact, diverse candidate set for matching, formalized by a scoring expression that balances ranking position and access time. Online experiments on Taobao show substantial gains, including up to 57% improvement in CTCVR and a 5% uplift in transactions, validating the approach's effectiveness and efficiency. The results imply that model-free, cross-scenario nearline retrieval can significantly enhance the matching stage's recall while meeting latency constraints, with strong potential for broader adoption across platforms.

Abstract

In recommendation systems, the matching stage is becoming increasingly critical, serving as the upper limit for the entire recommendation process. Recently, some studies have started to explore the use of multi-scenario information for recommendations, such as model-based and data-based approaches. However, the matching stage faces significant challenges due to the need for ultra-large-scale retrieval and meeting low latency requirements. As a result, the methods applied at this stage (collaborative filtering and two-tower models) are often designed to be lightweight, hindering the full utilization of extensive information. On the other hand, the ranking stage features the most sophisticated models with the strongest scoring capabilities, but due to the limited screen size of mobile devices, most of the ranked results may not gain exposure or be displayed. In this paper, we introduce an innovative multi-scenario nearline retrieval framework. It operates by harnessing ranking logs from various scenarios through Flink, allowing us to incorporate finely ranked results from other scenarios into our matching stage in near real-time. Besides, we propose a streaming scoring module, which selects a crucial subset from the candidate pool. Implemented on the "Guess You Like" (homepage of the Taobao APP), China's premier e-commerce platform, our method has shown substantial improvements-most notably, a 5% uptick in product transactions. Furthermore, the proposed approach is not only model-free but also highly efficient, suggesting it can be quickly implemented in diverse scenarios and demonstrate promising performance.

Figures (2)

• Figure 1: Different scenarios at Taobao.
• Figure 2: Multi-Scenarios Nearline Retrieval Framework on Taobao. As users engage with different scen arios, the fine-ranked outcomes are queued for processing. They are then scored and truncated via Flink, before being stored in an online database. Whenever users visit the homepage, the relevant data is fetched from the online database and fed into the matching process.