Beyond Existing Retrievals: Cross-Scenario Incremental Sample Learning Framework

TL;DR

The paper tackles diminishing returns from cross-scenario samples in large-scale multi-retrieval recommender systems by proposing IncRec, a framework that learns cross-scenario incremental samples not retrieved by any existing model. It constructs two sample groups (RTG and ITG) and employs a three-tower architecture with a basic dual-tower, an incremental user tower, and a consistency-aware alignment tower, optimized jointly to favor incrementally useful items. Key contributions include the explicit construction of extreme cross-scenario incremental samples, an incremental learning objective with adaptive consistency weighting, and a alignment mechanism to align with downstream ranking. Extensive offline benchmarks on KuaiRand and Taobao data, plus online A/B testing on Taobao, demonstrate substantial incremental gains in hit rate, exposure alignment, PVR, and transaction count, validating practical impact.

Abstract

The parallelized multi-retrieval architecture has been widely adopted in large-scale recommender systems for its computational efficiency and comprehensive coverage of user interests. Many retrieval methods typically integrate additional cross-scenario samples to enhance the overall performance ceiling. However, those model designs neglect the fact that a part of the cross-scenario samples have already been retrieved by existing models within a system, leading to diminishing marginal utility in delivering incremental performance gains. In this paper, we propose a novel retrieval framework IncRec, specifically for cross-scenario incremental sample learning. The innovations of IncRec can be highlighted as two aspects. Firstly, we construct extreme cross-scenario incremental samples that are not retrieved by any existing model. And we design an incremental sample learning framework which focuses on capturing incremental representation to improve the overall retrieval performance. Secondly, we introduce a consistency-aware alignment module to further make the model prefer incremental samples with high exposure probability. Extensive offline and online A/B tests validate the superiority of our framework over state-of-the-art retrieval methods. In particular, we deploy IncRec in the Taobao homepage recommendation, achieving a 1% increase in online transaction count, demonstrating its practical applicability.

Figures (2)

• Figure 1: Cross-scenario incremental sample construction
• Figure 2: Architecture of IncRec. The model employs a shared item tower and three specialized user towers: the basic tower captures overall user interests, the incremental tower focuses on incremental interests, and the alignment tower ensures consistency with downstream ranking preference.