How Alipay’s Homepage Leverages Edge AI for Smarter Refreshes

1. Introduction

The Alipay app homepage is the first tab page, responsible for traffic distribution and entry experience. By deeply cooperating with the edge‑intelligence team, the homepage leverages edge computing and real‑time client behavior features combined with algorithm platform capabilities to boost efficiency and value.

2. Homepage Moving Toward Intelligence

2.1 Background

Homepage recommendation has matured; the technology must upgrade to support business, digitization, and future needs. For recommendation zones, product‑side requires scenario‑based recommendation and fine‑grained operation, while technical‑side demands high‑efficiency integration and personalized configuration, making core capability expansion essential.

2.2 Intelligent Goals

The homepage must meet product stability, flexibility, and efficiency, as well as technical performance, stability, experience, and digitization requirements. Intelligent construction aims to raise the business baseline while continuously deepening technology to ensure product experience.

3. Intelligent Scenarios on the Homepage

3.1 Smart Refresh

Refresh is the core capability of the distribution platform. Previously, refresh strategies covered all possible scenarios. Now they are abstracted into three dimensions—time, space, and event—and each is upgraded intelligently.

3.1.1 Time Dimension

The current cloud‑recommend‑client‑render architecture uses coarse thresholds, leading to unreliable timing, lack of continuity, and single‑mode strategies.

Unreliable time thresholds – high frequency wastes resources, low frequency reduces effective reach.

Partial lack of continuity – cloud recommendations rely only on prior behavior and historical profiles.

Single strategy mode – adjustments depend on large version or policy changes, limiting personalization.

To address these, an AB experiment with gradient thresholds and client‑side models is introduced, enabling real‑time intent recognition and building an edge‑cloud collaborative recommendation architecture.

3.1.2 Space Dimension

Near‑field recommendation uses precise location to deliver services in places like stations, airports, hospitals, etc. Challenges include over‑reliance on refresh thresholds and non‑real‑time location perception.

Solution: a combination of “coverage circle + geofence buffer” that dynamically adjusts client‑side location frequency based on distance to areas of interest, reducing unnecessary requests and improving service efficiency.

Position 1 – server defines a coverage radius R; client uses low/medium frequency within the circle.

Position 2 – client checks movement; if still inside, keep current state.

Position 3 – if leaving the circle, request an updated coverage.

Position 4 – entering the geofence buffer triggers higher‑frequency location checks.

Position 5 – when inside the buffer, request service cards; exit triggers card reclamation.

3.1.3 Event Dimension

Many refresh scenarios depend on user events (e.g., pull‑to‑refresh, feedback). Existing mechanisms suffer from weak event awareness and low‑efficiency asynchronous flows.

Weak event awareness – relies on external notifications; complex events lack direct integration.

Low async efficiency – only ~10% of async notifications succeed, wasting resources.

The team redesigned the event‑driven refresh mechanism, creating a perception‑based intelligent refresh solution and lightweight business integration, dramatically improving efficiency.

3.2 Recommendation Deduplication

Duplicate recommendations arise from multiple traffic slots (search, badges, feeds). The solution introduces centralized traffic control both on the cloud (Single‑Request algorithm) and on the client (edge‑intelligence algorithm).

Solution 1 – Single‑Request Cloud Deduplication : merges requests on the cloud, using Recmixer as the main link and UCDP for coordination.

Pros: joint optimization across slots.

Cons: high latency, heavy client changes.

Solution 2 – Edge‑Intelligence Deduplication : implements a client‑side control platform caching recommendations and making unified decisions.

Pros: low latency, decoupled systems.

Cons: greedy first‑come‑first‑served algorithm.

After ROI comparison, the edge‑intelligence approach was adopted, combining offline labeling and online rule‑based deduplication.

4. Phase Progress and Results

Smart Refresh – real‑time intent prediction reduces request volume by 14%, increases click PV confidence by 0.44%, and lifts overall PVCTR by 18.8%.

Spatial Intelligence – geofence buffer improves click PV/UV by 19.7%/20.02% and PVCTR/UVCTR by 4.4%/6.51%; coverage circle boosts PV/UV by 4.54%/2.54% and PVCTR/UVCTR by 2.38%/0.51%.

Event Intelligence – business integration efficiency greatly improves; async sync requests drop by 90%.

Recommendation Deduplication – online de‑duplication for waistband, feed, and badge slots reduces duplicate entities by 33%.

5. Future Plans

Continue deepening the smart refresh architecture, expanding edge‑cloud collaboration, and focusing on:

Deeper content entity mining – enhance edge‑side granularity to improve conversion.

Broader edge decision scope – turn invalid refreshes into effective requests with personalized algorithmic decisions.

Further work includes designing intelligent detection for user experience assurance, building online monitoring, anomaly alerts, and issue discovery mechanisms.