How Edge Intelligence Revolutionizes Real-Time Recommendations on Mobile Apps

Background

Information flow is a major traffic entry for Taobao, crucial for browsing efficiency, conversion, and traffic distribution. To recommend products users like, Alibaba first applied large‑scale edge computing on the Taobao client, using rich user features and triggers on the device, together with machine learning and deep neural networks, to continuously sense user intent and provide timely feedback.

After half a year of continuous improvement, edge intelligence for Taobao’s information flow was fully rolled out in mid‑September, delivering a substantial increase in click‑through rate and GMV on Double 11.

Current Situation and Solution

In list‑based recommendation scenarios such as Taobao’s information flow, users often browse without a clear purchase intent, discovering items gradually. Traditional cloud recommendation systems trigger sorting on the server after a client request and then present the sorted items, which suffers from two main problems:

Limited opportunities to adjust content on the client side, usually only at pagination requests.

Inability to capture the user’s current preference instantly, leading to delayed feedback.

These issues cause a mismatch between user preference changes and the system’s perception, reducing click and conversion rates.

Edge Intelligence Changes

Edge intelligence combines “edge” (device‑side data and computation) with “intelligence” (machine‑learning‑driven decision making). By integrating rich user features and real‑time computation on the device, the system can sense intent instantly, make decisions locally, and provide immediate feedback, overcoming the latency and adjustability limitations of cloud‑only solutions.

Edge vs Cloud Comparison

Advantages of Edge Intelligence

Abundant user features and touchpoints on the device enable more informed decisions.

High real‑time performance saves network transmission time and speeds up feedback.

Utilizes device compute and storage, reducing cloud resource consumption.

All data processing stays on the device, enhancing privacy.

Disadvantages of Edge Intelligence

Device resources (compute, power, storage) are limited, restricting large‑scale continuous computation.

Algorithm scale is smaller; models cannot be as large as cloud counterparts.

Device‑side data is limited in volume and longevity.

Advantages of Cloud Intelligence

Access to massive, long‑term data for big‑data analytics.

Scalable compute, power, and storage resources.

Large‑scale models can achieve optimal solutions.

Disadvantages of Cloud Intelligence

Higher latency due to network transmission.

Weaker real‑time perception of user intent because of data transfer delays.

The optimal approach is a collaborative cloud‑edge intelligence model, leveraging the strengths of both.

Edge Intelligence Infrastructure

The system is built on five core components:

Data – BehaviX : A device‑side user feature data center that provides real‑time synchronized features to the cloud, enabling intent analysis.

Decision Framework – BehaviR : Simplifies business integration, allowing real‑time perception and timely intervention.

Edge Computing – EdgeRec : Runs on‑device logic or models, performing real‑time user state modeling and supporting multi‑task learning.

Computation Engine – Walle & MNN : Provides a unified runtime (Walle) and a lightweight deep‑learning inference engine (MNN) for both iOS and Android.

Algorithm Platform – Jarvis : Offers end‑to‑end development, debugging, A/B testing, deployment, and monitoring, supporting feature and sample computation libraries.

These components are orchestrated by an on‑device scheduling system that connects user perception with decision making, while the cloud supplies long‑term features and content.

Data Effect

Since the launch of edge intelligence in the information‑flow “Guess You Like” scenario, continuous experiments have steadily improved performance. On Double 11, the system achieved a notable increase in both click‑through rate and GMV, demonstrating the impact of real‑time, personalized recommendations.

Conclusion

When the device can not only present content but also “think,” it transcends the limitations of cloud‑only solutions. Edge intelligence enables low‑latency decision making, reduces resource consumption, and opens new possibilities for integrating sensor data, UI interactions, and personalized algorithms, paving the way for more innovative, real‑time user experiences.