How Alibaba’s CLOES Algorithm Cut Search Latency by 30% During Double‑11

Background

On June 29, 2017 Alibaba held the Tmall Double‑11 launch event and announced the “Super IP Plan”. The same event also revealed the research results that supported the 2016 Double‑11 shopping festival.

Paper Overview

The paper “Cascade Ranking for Operational E‑commerce Search” was accepted at KDD 2017. It proposes a cascade‑style ranking method (CLOES) that splits a single ranking task into multiple stages, using increasingly complex features at later stages.

Key Challenges

Taobao’s search engine faces massive traffic, especially during large‑scale promotions like Double‑11, leading to high CPU load and latency. Moreover, e‑commerce ranking must optimize multiple objectives (click‑through rate, transaction volume, GMV) under strict latency and result‑size constraints.

Limitations of Existing Methods

Most learning‑to‑rank research focuses on ranking quality while ignoring efficiency. Traditional two‑stage ranking (simple features → candidate set → complex ranking) does not guarantee optimal trade‑offs between performance and effectiveness.

CLOES Algorithm

Inspired by cascade object detection, CLOES treats each item as a k‑dimensional vector and estimates the probability of passing each stage using a sigmoid function. The overall click probability is the product of stage‑wise probabilities.

The loss function combines a ranking accuracy term (negative log‑likelihood) with constraints on latency (<100 ms) and minimum result count (≥200), formulated similarly to an SVM‑style loss.

Additional multi‑objective factors (click, transaction volume, GMV) are incorporated to balance performance in the e‑commerce scenario.

Experimental Results

Offline experiments confirmed the algorithm’s effectiveness. Online A/B tests during the 2016 Double‑11 promotion showed CPU usage dropping from 32 % to 18 % and average search latency decreasing from 33 ms to 24 ms, while CTR and GMV slightly increased.

On the peak of Double‑11, the engine load never exceeded 70 %, CPU usage was reduced by about 45 %, and average latency fell by roughly 30 %, contributing to an estimated 1 % GMV increase. The more efficient ranking also enabled the deployment of additional costly features such as real‑time and RNN‑based signals, further boosting CTR by 10‑20 %.

Conclusion

Search is the primary traffic source for e‑commerce platforms, and high‑quality ranking directly impacts user experience, merchant revenue, and system efficiency. CLOES demonstrates that a carefully designed cascade ranking system can simultaneously improve effectiveness and efficiency, setting a foundation for future advances in Alibaba’s search infrastructure.