COTTI: Contrastive Training and Time-Aware Inference for Robust Industrial Recommendation under Promotional Drift

Maolei Huang, Jing Tian, Dexuan Duan, Fangfang Lin, Huawei Cao, Shuhan SongarXiv 2026 · arXiv

TL;DR

COTTI is a robust recommendation framework designed to handle "promotional drift"—the sudden shift in user behavior during major sales events like Black Friday or Double 11. By combining adaptive contrastive training with an uncertainty-aware inference adjustment, it achieves significant gains in conversion rates and revenue while maintaining the low latency required for industrial-scale deployment.

What Problem Does This Paper Address?

In the world of e-commerce, promotional events are both a blessing and a curse for recommendation systems. During a sale, traffic patterns, purchase intent, and conversion scales shift dramatically. A user who typically browses budget flights might suddenly book a luxury hotel package because of a limited-time discount.

Standard recommendation models face a dilemma: if they are too stable, they react too slowly to these transient "peaks"; if they adapt too aggressively, they might lose track of a user's long-term preferences or get distracted by low-intent "window shopping." This phenomenon is known as promotional drift. The authors of COTTI set out to build a system that maintains high-quality rankings and reliable conversion-rate (CVR) estimates even when the ground shifts beneath the model's feet.

Methods

The COTTI (COntrastive Training and Time-aware Inference) framework tackles the problem at three different levels:

  1. Cross-Period Contrastive Alignment (CPC): During training, this component encourages a user’s "regular" history and "promotional" history to share information. Crucially, it uses an adaptive consistency weight—it doesn't force the two views to be identical if the user's behavior has legitimately changed, allowing for useful "discovery" of new interests.
  2. Active Re-weighting (RW): Not every interaction during a sale is equally meaningful. RW calculates the "distance" between a promotional interaction and a user's historical interests. It then downweights examples that look like random browsing while preserving the influence of high-intent actions.
  3. Adaptive Boundary Alignment (ABA): This is a clever serving-time trick. Once the model is trained and its parameters are fixed, ABA applies a "logit correction" during inference. It uses MC dropout to measure predictive uncertainty (how "unsure" the model is about a specific user-item pair) and combines this with a temporal kernel that tracks the intensity of the promotional event. If the model is uncertain during the peak of a sale, ABA gives the prediction a principled "nudge" upward.
Fig. 1. COTTI workflow. CPC and RW shape the predictor; earlier validation selects calibration settings; ABA adjusts scores with fixed predictor parameters during serving.

Datasets

The authors evaluated COTTI using massive, real-world industrial datasets from Fliggy (Alibaba’s travel platform). The data includes multi-month logs covering major Chinese shopping festivals:

  • 6.18 Mid-year Sale
  • Double 11 (Singles' Day)
  • Double 12

The offline evaluation used over 100 million samples, providing a rigorous testbed for handling extreme distribution shifts.

Key Insights and Ideas

  • Separation of Concerns: One of the most powerful ideas in COTTI is separating representation learning (how we understand the user) from score adjustment (how we react to the event). This allows the model to stay "smart" about users while being "aware" of the calendar.
  • Uncertainty as a Correction Signal: Using predictive variability (via dropout) as a proxy for under-prediction is an elegant way to handle calibration. Instead of a flat bias for everyone during a sale, COTTI applies corrections where the model is most likely to be "confused" by the drift.
  • The Discovery Trade-off: The paper highlights a fascinating tension: better overall prediction accuracy (AUC) can actually lead to lower recall for "novel" categories. In other words, making the model more robust can sometimes make it more conservative, a critical insight for engineers balancing short-term revenue with long-term user exploration.
Schematic of the three mechanisms: CPC aligns views, RW weights by distance, and ABA modifies scores based on event timing and uncertainty.

Key Results

COTTI demonstrated superior performance in both offline and online settings:

  • Offline Performance: It achieved an AUC of 0.8127, outperforming the previous state-of-the-art (TEN) at 0.8101.
  • Online Impact: In a 10-day A/B test on Fliggy, COTTI delivered a 3.14% relative gain in conversion rate (CVR) and a 2.68% increase in revenue per mille (RPM) compared to the existing production model.
  • Efficiency: Despite the complexity of running multiple stochastic passes for ABA, the system maintains a p99 latency of 38.9 ms, which is well within the requirements for high-traffic industrial environments.
Online CVR and RPM lift during full deployment, showing consistent gains across different times of the day.

Limitations and Future Work

While COTTI is highly effective, the authors note a few areas for improvement:

  • Novelty Recall: As mentioned, the framework can sometimes suppress "novel category" discovery. Future work could look at ways to protect these exploratory interactions.
  • Event Scheduling: The ABA component relies on a pre-defined "event kernel" based on announced campaign timings. A more autonomous system might learn to detect these shifts in real-time without manual scheduling.
  • Computational Overhead: Although the latency is production-ready, running 10 stochastic passes (M=10) for every request is computationally expensive. Optimizing this "stochastic path" remains an open engineering challenge.

Summary updated

Abstract

Promotional events shift traffic composition, purchase behavior, and the scale of conversion probabilities on different time horizons. We present \textbf{COTTI} (\textbf{CO}ntrastive \textbf{T}raining and \textbf{T}ime-aware \textbf{I}nference), an industrial recommendation framework that separates historical representation learning from event-time score adjustment. Cross-Period Contrastive Alignment (CPC) regularizes paired histories of the same user with adaptive consistency weights. Active Re-weighting (RW) controls the influence of promotional examples through historical-interest distance. Adaptive Boundary Alignment (ABA) combines stochastic prediction variability with a temporal event kernel to adjust logits while keeping predictor parameters fixed. We analyze the weighted objective and bound ABA-induced probability and ranking changes. On Fliggy data, COTTI attains an offline AUC of 0.8127, compared with 0.8101 for TEN and 0.8123 for TEN with ABA. Cross-category analysis reveals improved discrimination alongside declines in paid-item recall for novel categories. End-to-end measurements give a p99 latency of 38.9\,ms and throughput of 14,700 requests/s; enabling ABA adds 2.2\,ms to p99. A ten-day online experiment against the serving HDR model yields relative gains of 3.14\% in conversion rate and 2.68\% in revenue per mille. Together, these results characterize the trade-off between aggregate prediction, novel-category discovery, and serving cost in promotional recommendation.
BibTeX
@misc{huang2026cotti,
  author = {Maolei Huang and Jing Tian and Dexuan Duan and Fangfang Lin and Huawei Cao and Shuhan Song},
  title = {{COTTI: Contrastive Training and Time-Aware Inference for Robust Industrial Recommendation under Promotional Drift}},
  year = {2026},
  eprint = {2611.22597},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.22597},
}
Figure