Custom Model-Powered AIRec Recommendations

Scenario

Use this workflow when you need to replace AIRec’s default ranking algorithms with a proprietary or fine-tuned model (e.g., a deep learning CTR predictor) while retaining AIRec’s orchestration, real-time user targeting, and A/B testing capabilities. PAI-EAS serves the custom model with GPU/CPU auto-scaling, and AIRec routes live traffic to the endpoint for personalized item scoring.

Integration steps

1. Package & Upload Model: Export your trained model (SavedModel/ONNX) and upload to an OSS bucket. Bundle preprocessing/postprocessing scripts if deploying a full pipeline.
2. Deploy to PAI-EAS: Use the PAI console or OpenAPI to create an Elastic Algorithm Service.

``bash POST /api/v2/services { "name": "custom-ranker-v1", "model_path": "oss://<bucket>/model.tar.gz", "processor": "tensorflow", "auto_scaling": {"min_instances": 2, "max_instances": 10, "target_cpu": 70} } ``

3. Verify Endpoint: Confirm the service reaches Running state and capture the VPC/public URL (e.g., https://<service-id>.<region>.pai-eas.aliyuncs.com).
4. Deploy AIRec Instance: If your target region is cn-hangzhou or another supported zone, deploy via console or API:

``bash POST /v2/instances {"region": "cn-hangzhou", "service_type": "custom_ranking", "vpc_id": "<vpc-id>"} ``

5. Bind PAI-EAS to AIRec: In AIRec console under Algorithm > Custom Ranking, or via UpdateInstanceConfig:

``json {"custom_ranker": {"endpoint": "https://<service-id>.pai-eas.aliyuncs.com", "timeout_ms": 500, "request_format": "json"}} ``

6. Ingest Data: Push user, item, and behavior logs using AIRec’s BatchImport API to populate the feature store.
7. Configure A/B Testing: Create an experiment via POST /v2/experiments splitting traffic 50/50 between the default AIRec ranker and your PAI-backed model.
8. Validate: Call AIRec’s Recommend API and verify x-pai-eas-request-id appears in response headers, confirming successful routing.

Architecture

Client requests hit the AIRec Gateway, which fetches real-time user/item features from the AIRec Feature Store. AIRec forwards feature payloads to the PAI-EAS endpoint for custom scoring. PAI-EAS returns ranked item scores, which AIRec applies business rules, diversity filters, and A/B routing logic to before returning the final recommendation payload. PAI manages inference compute and auto-scaling; AIRec manages traffic orchestration and experimentation.

Prerequisites

• Alibaba Cloud account with PAI and AIRec activated
• Trained model artifact (SavedModel/ONNX) stored in OSS
• PAI-EAS service role with AliyunPAIEASDefaultRole and VPC network access
• AIRec instance quota in a supported region (e.g., cn-hangzhou)
• VPC peering or public endpoint configured between AIRec and PAI-EAS

Common pitfalls

• Network Isolation: AIRec and PAI-EAS in separate VPCs cause timeouts. Use VPC peering or expose PAI-EAS publicly with AIRec IP whitelisting.
• Payload Format Mismatch: AIRec sends JSON feature arrays, but PAI-EAS defaults to Protobuf. Explicitly set "request_format": "json" in the EAS service spec.
• Cold Start Latency: PAI-EAS scaling cold starts exceed AIRec’s 500ms timeout. Set min_instances: 2 and enable model warm-up.
• Region Constraint Violation: AIRec deployment is restricted to specific regions. Deploy PAI-EAS in the exact same region as the AIRec instance to avoid cross-region routing failures.

Typical questions

• deploy custom recommendation model
• PAI model for AIRec
• custom ranking model with recommendations
• deploy model behind AIRec
• personalized recommendations with custom model
• deploy model and serve recommendations
• PAI inference integrated with AIRec
• 部署自定义推荐模型

Q: How do I integrate a custom model with AIRec for personalized recommendations? A: You deploy a custom ranking or scoring model on PAI-EAS for online inference and integrate it with AIRec as the personalization backend. PAI handles model serving with auto-scaling, while AIRec orchestrates recommendation delivery, A/B experimentation, and real-time user targeting.