How LAST Enables Real‑Time Learning for Re‑Ranking in E‑Commerce Recommendations

Background: Real-time Online Learning & Re-ranking Model Basics

Introduce concepts of real-time online learning and re-ranking models in industrial recommendation pipelines. Traditional offline training takes days; online learning aims to update models instantly using new samples.

Challenges of Real-time Online Learning

Dependence on user feedback latency (hours to days).

Need for real-time data streaming infrastructure.

Model stability concerns.

Re-ranking Model Overview

Re-ranking is the final stage of a typical recommendation pipeline, refining candidate items (recall → coarse → fine → re-rank). It explicitly models context interactions among items.

Two modeling streams: pointwise scoring with context, and sequence generation without explicit rank scores.

Formal Introduction: Definition & Solutions

Define the re-ranking problem as selecting the optimal ordered subset of candidates (e.g., three candidates to output two items, six possible permutations). The objective is to maximize a reward function reflecting user, platform, and merchant feedback.

New Proposal: Learning at Serving Time (LAST)

LAST updates the model at the moment a request arrives, without waiting for user feedback. It computes a per‑request parameter shift Δθ to improve the immediate prediction, then discards it after serving.

Architecture

Classic online learning updates the deployed model θ after receiving feedback samples. LAST keeps the deployed model fixed and, for each request (u, c), searches for Δθ that maximizes the evaluator score.

Solution Steps

One‑shot solution: call a generator to produce a sequence, evaluate it, and adjust generation probabilities.

Introduce an evaluator that predicts feedback for alternative sequences, enabling training without real feedback.

Parallel version: generate multiple Δθ candidates via gradient exploration, evaluate each, and select the best sequence.

Experimental Results

Offline evaluation using NDCG shows that generator‑evaluator multi‑sequence methods outperform baselines, with LAST achieving the highest score. A second evaluator‑based offline test confirms a >1‑point gain over the strongest baseline.

Online A/B test in Taobao’s “follow” feed shows that LAST increases transaction volume by two points while keeping click volume stable.

Conclusion

LAST provides real‑time model updates without feedback latency, offers per‑request optimization, is plug‑in compatible, requires no additional online infrastructure, and has been validated both offline and online.

The work was a collaboration between Alibaba and Renmin University, involving researchers Wang Yuan, Li Zhiyu, Wen Zijian, Lin Quan, Zhang Changshuo, Chen Sirui, Zhang Xiao, and Xu Jun.