How Alibaba’s ‘Ali Xiaomi’ Prediction Platform Boosts Smart Customer Service with AI

Background

Ali Xiaomi, launched in 2015, is Alibaba’s intelligent customer‑service robot. By 2017 it handled 6.43 million requests during Double 11, achieving a 95% intelligent resolution rate. To further reduce the pressure on smart‑customer‑service bots, Alibaba initiated a prediction platform that forecasts user intent before interaction, enabling proactive or reactive assistance.

Core Capabilities

The platform focuses on two main predictions:

Order Prediction : predicts which user order may encounter a problem.

Issue Prediction : predicts the type of problem the user is facing.

Additional capabilities include scenario prediction (e.g., account, logistics, rights), prior‑knowledge pattern prediction, and user prediction (identifying the Taobao ID behind a phone call).

Deployed Products

The platform now serves multiple Alibaba chatbot families—Ali Xiaomi, Store Xiaomi, Hotline Xiaomi, XSpace, and Wanxiang—delivering the following results:

Order prediction improves click‑through rates by 16‑20% and raises satisfaction by 3.2%.

Issue‑prediction “guess you want to ask” mode boosts click‑through by 2‑3×; a high‑accuracy mode raises click‑through by nearly 10×.

Issue‑prediction bot increases effective rate by 35 percentage points and satisfaction by 3 points, though coverage is around 3%.

Hotline Xiaomi achieves 90% correct predictions and a 7% satisfaction lift.

Store Xiaomi’s recent rollout of after‑sale knowledge prediction raises click‑through and resolution rates by 4% each.

XSpace’s order prediction reduces manual order‑number entry, cutting ATT by over 6% and saving millions of RMB annually.

Wanxiang’s chatbot improves click‑through and resolution rates by 4% in consumer‑protection and transaction scenarios.

Algorithmic Foundations

The platform’s algorithm stack includes deep‑CTR models such as DeepFM, PNN (IPNN), and DCN. DCN extends Wide & Deep by efficiently learning sparse and dense feature interactions with fewer parameters. The system also experiments with label‑LDA, XGBoost, and other classifiers.

Recent enhancements to the DeepCTR suite comprise:

Input layers supporting both one‑hot and multi‑hot features.

Independent training of word vectors using the 2‑vect algorithm.

Incorporation of text‑embedding layers.

Reinforcement Learning

To further improve click‑through, solve‑rate, and satisfaction, the team combines the deep‑CTR base model with a DRL‑based reranking layer that uses sequential episode modeling based on real‑time feedback.

Streaming Computation

Beyond reactive prediction, the platform leverages stream processing to monitor user logs in real time, proactively detecting issues and pushing timely assistance. This is realized via Blink and EAS services that ingest live browsing logs and output predictions for immediate user outreach.

Platform Architecture

The solution is organized into four layers:

Execution Engine : orchestrates prediction workflows via configurable components, replacing hard‑coded pipelines.

Offline Computing : aggregates data from ODPS, TT logs, and TC APIs for feature generation and model training.

Real‑Time Computing : processes live logs to produce instant predictions, feeding both direct user messages and downstream post‑interaction models.

Management Backend : provides operations, flow management, automated model rollout, gray‑scale control, and service degradation.

The platform adopts a modular design:

Component : the smallest unit implementing a specific algorithm or data fetch.

Module : an ordered sequence of Components.

Switch : controls flow logic (conditions, parallelism, loops) linking Modules.

Flow : a DAG of Modules and Switches that realizes a complete prediction service.

Biz : a business‑level configuration composed of multiple Flows, identified by a biz‑id.

Conclusion

The prediction platform has become a core enabler for Alibaba’s chatbot ecosystem, delivering measurable improvements in user experience while streamlining code reuse, model evaluation, and service management. Ongoing work continues to expand coverage, refine algorithms, and enhance real‑time proactive assistance.