ConAlign: Conditional Alignment Framework for Balancing Biased and Unbiased Recommendation
Jingcheng Zhang, Yihan Wang, Qi Song, Liyin Hong
cs.IR
2026-07-27
ConAlign's conditional alignment between biased and unbiased recommendation towers improves 7-day retention by 0.029% and interest diversity by 0.097% in Kuaishou production A/B testing, while running 86× faster than the prior state-of-the-art InterD and achieving unbiased UAUC of 0.6841 on Yahoo! R3.
Recommender systems trained on logged user behavior absorb exposure bias (only shown items receive interactions), selection bias (popular items have naturally higher engagement), and position bias. The model learns the platform's display policy rather than genuine user preferences, reinforcing popular items and narrowing the content distribution. Kuaishou's team empirically observed that user interest diversity (Valid Interest Number, VIN) correlates positively with 7-day retention (LT7).
The standard debiasing approach introduces random-traffic data: items inserted at random positions with a small fixed probability, producing approximately missing-at-random (MAR) interactions. Two obstacles remain: random data volume is tiny; aggressive debiasing can hurt accuracy in the biased serving environment, degrading short-term engagement. The prior state-of-the-art InterD uses bi-level meta-learning and full Cartesian-product imputation, requiring 10,592 seconds to train on Yahoo! R3 — unusable in a streaming online system.
ConAlign uses a dual-tower architecture. The biased tower trains on large-scale logged data with full feature sets (memorization features capturing historical co-occurrence plus generalization features). The unbiased tower trains on the small random-traffic stream using only generalization features.
The key innovation is the conditional alignment loss. The unbiased tower's total loss is Lu = Ld + λ × Lalign, where Lalign = Icond × (1/N) Σ ‖hd − sg(hb)‖. The gate Icond = 1 if and only if the biased tower's per-instance prediction loss on that sample is lower than the unbiased tower's. sg(·) stops gradients from propagating to the biased tower.
Alignment only transfers knowledge on samples where the biased tower is locally more reliable. On samples where the biased tower fails, no alignment occurs and the unbiased tower learns from its own MAR signal. Alignment operates at the final hidden-layer representation level, not the prediction score level, preserving richer structural information. At inference, only the unbiased tower's score is used.
Random traffic is collected via a permanently running intervention mechanism: at a small fixed probability, one item sampled uniformly from the full catalog is inserted at a random feed position with negligible user experience impact.
Offline (Coat, Yahoo! R3, KuaiRand-Pure, evaluated on unbiased test sets):
| Dataset | ConAlign UAUC | InterD (2nd best) | Gain |
| Coat | 0.6856 | 0.6792 | +0.0064 |
| Yahoo! R3 | 0.6841 | 0.6787 | +0.0054 |
| KuaiRand-Pure | 0.6372 | 0.6251 | +0.0121 |
Training time on Yahoo! R3: ConAlign 122.7s vs. InterD 10,592.1s — approximately 86× faster.
Kuaishou production A/B test (10% of user traffic, long-term observation after 7-day AA confirmation):
| Metric | Change | Significant |
| DAU | +0.069% | Yes |
| 7-day retention (LT7) | +0.029% | Yes |
| Interest diversity (VIN) | +0.097% | Yes |
| Content concentration (CC, lower = more diverse) | −0.083% | Yes |
| Completion plays | +0.277% | Yes |
DAU and LT7 improvements at this scale on a mature recommendation system are practically significant.
ConAlign turns a well-known dilemma (debiasing versus factual accuracy) into an engineering choice. The conditional gate is the key: knowledge flows from the biased tower to the unbiased tower only where the biased tower is locally more reliable, avoiding over-regularization on samples where the unbiased signal should dominate. This explains why ConAlign simultaneously improves unbiased test accuracy and biased-environment online metrics. The 86× speed advantage makes it viable in streaming production training.
The framework requires a permanently running random-traffic intervention infrastructure; platforms without this cannot directly apply it. MAR validity depends on truly unbiased insertion; systematic position selection breaks theoretical guarantees. The alignment weight λ requires careful tuning, as excessive values significantly hurt unbiased UAUC. Offline experiments use MF as a unified backbone for fair comparison; the gap to production-grade DNNs is acknowledged but not closed.