Abstract
Industrial recommendation systems often incorporate cross-scenario data (e.g., using "Search" behaviors to improve "Homepage" results) to expand the retrieval candidate pool. However, standard methods frequently suffer from redundancy, as many cross-scenario items are already captured by existing retrieval models. This paper introduces IncRec, a framework that specifically targets "incremental" samples—items that current models fail to retrieve. By utilizing a multi-tower architecture that includes a specialized alignment tower to match downstream ranking preferences, IncRec significantly improves retrieval efficiency and conversion rates in large-scale e-commerce environments.
Introduction: The Challenge of Redundancy
Most large-scale recommendation systems operate in two stages: Retrieval, which narrows millions of items down to a few thousand candidates, and Ranking, which performs fine-grained scoring. To improve the retrieval "ceiling," engineers often pull in data from auxiliary scenarios. For instance, if a user purchases an item via "Search," that item becomes a strong candidate for the "Homepage" recommendation.
The authors identify a major inefficiency in this approach: Redundancy. Many cross-scenario items are already being found by existing retrieval methods like Collaborative Filtering or Graph-based models. Simply adding more of the same data leads to diminishing marginal utility. The system expends computational resources relearning known patterns rather than discovering "incremental" items that could drive new conversions.
The IncRec Framework
IncRec addresses this by focusing on the "delta"—the items the current system misses. The framework is built on three core components:
1. Incremental Sample Construction
Cross-scenario data is partitioned into two distinct groups:
- Retrieved Target Group (RTG): Items already successfully captured by the existing retrieval ensemble.
- Incremental Target Group (ITG): "Extreme" samples that are preferred by users in other scenarios but are currently invisible to the homepage retrieval models.
2. Multi-Tower Architecture
To process these samples, IncRec employs a sophisticated multi-tower structure:
- Basic User Tower: Learns from all available data to ensure system stability.
- Incremental User Tower: Uses contrastive learning to specifically distinguish ITG items from RTG items. By treating ITG as positives and RTG as negatives, the model is forced to learn the unique features of these previously "unfound" items.
- Alignment Tower: Acts as a bridge to the Ranking stage. It ensures that the new items found by the retrieval stage won't be immediately discarded by the Ranker.
3. Consistency-aware Alignment
A common failure in retrieval is the "mismatch" with the ranker. If a retrieval model finds a relevant item that the Ranker hasn't seen during training, the Ranker may score it poorly. IncRec’s alignment tower uses an adaptive weighting mechanism based on representation similarity to prioritize incremental items that align with the Ranker's historical "exposed" preferences.
Experimental Results
The framework was evaluated on the KuaiRand video dataset and a massive Taobao industrial benchmark (21 days of logs across Homepage, Search, Shop, etc.).
- Offline Gains: IncRec-enhanced models (YouTube-DNN, DSSM, SASRec) showed significant improvements in "Incremental Hitrate," proving the model successfully targets the system's blind spots.
- Online Impact: In live A/B testing on the Taobao homepage, IncRec delivered a 1% increase in transaction count and a 9.07% increase in hitrate.
- Efficiency: The model achieved a 33% higher Proportional Visibility Rate (PVR), indicating that the items it retrieved were highly compatible with the downstream ranking stage.
Conclusion
IncRec demonstrates that in mature recommendation systems, the key to further growth lies in identifying and learning from incremental samples. By explicitly modeling the "unfound" and aligning retrieval with ranking preferences, the framework provides a scalable way to break through performance plateaus in multi-scenario environments. Future work may explore how to adapt this framework for "cold start" users who lack extensive cross-scenario histories.
Summary updated