← News

ResearchSources: arXiv · 18 Nov 2026

COTTI: Alibaba-linked framework targets recommendation drift during promotional events

A paper on arXiv describes COTTI, a recommendation framework that separates historical user modeling from event-time score adjustment, reporting offline and online gains on Fliggy data alongside a novel-category recall trade-off.

In plain words

Online shops must keep recommending relevant items even when big sales events change what users click and buy. COTTI is a proposed system that learns a user's long-term interests from history, then adjusts scores at event time without retraining the model. Its authors report better conversion and revenue in a live test, but weaker recall for brand-new product categories.

What was announced

The paper, posted to arXiv by the COTTI authors, addresses promotional drift: the way sales events shift traffic composition, purchase behavior and conversion-probability scales across different time horizons. The framework splits the problem in two. Historical representation learning is regularized by Cross-Period Contrastive Alignment, which pairs the same user's histories across periods with adaptive consistency weights, while Active Re-weighting controls how much promotional examples influence training based on historical-interest distance. At inference, Adaptive Boundary Alignment combines stochastic prediction variability with a temporal event kernel to adjust logits while keeping the predictor's parameters fixed; the authors also bound the probability and ranking changes this introduces.

Reported results come from Fliggy data. Offline, COTTI reaches an AUC of 0.8127 versus 0.8101 for TEN and 0.8123 for TEN with ABA. In a ten-day online experiment against the serving HDR model, the authors report relative gains of 3.14% in conversion rate and 2.68% in revenue per mille. Serving measurements show p99 latency of 38.9 ms and throughput of 14,700 requests per second, with ABA adding 2.2 ms to p99. The authors note improved discrimination overall but declines in paid-item recall for novel categories, framing the results as a trade-off between aggregate prediction, novel-category discovery and serving cost.

What it does not show

  • Evidence is limited to the paper's abstract; the full methodology, baselines and experimental setup could not be checked.
  • All figures are author-reported on Fliggy data; no independent confirmation of the offline or online results is available.
  • The abstract does not name the authors' affiliations or state whether COTTI is deployed beyond the reported experiment.

For security and IT teams

Status
COTTI is described in a new arXiv paper with a ten-day online experiment against a serving HDR model; the abstract does not state general availability.

Sources

Source-support review completed 19 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.

Coverage: arXiv: abstract.

Updates and corrections

No corrections so far.

First published 19 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.