Usage notes
Before integrating PAI-Rec into your system, complete the following steps to confirm that you have the data, scenarios, and cloud services in place. Skipping these steps may affect service testing.
If you are building a recommendation system for the first time and lack recommendation algorithms or ML engineering experience, use Artificial Intelligence Recommendation (AIRec) instead — it is an end-to-end recommendation service that requires no custom modeling. The steps below apply only if you have the technical resources to independently control the recommendation system and want to use PAI-Rec.
Step 1: Collect behavioral data
A recommendation model requires rich user and item features, plus user behavioral signals. Before integration, confirm that your event tracking pipeline captures at minimum the following behaviors from the recommendation page:
Exposure — the item was shown to the user
Click — the user clicked the item
Consumption — downstream actions such as purchase, favorite, comment, share, like, order, or review
These three event types are the minimum required to train and calibrate recommendation models. To improve recommendation quality over time, also collect end-to-end behavioral signals such as search queries and activity participation.
If you are unsure about the required data formats, refer to the demo data.
Step 2: Define recommendation scenarios and success metrics
Before integration, confirm the following for each recommendation placement:
Pages — which pages will show personalized recommendations
Key metrics — the primary business outcomes attributed to recommendations
Auxiliary metrics — secondary signals used to monitor recommendation health
Attribution criteria — how you define whether a conversion was driven by a recommendation
Defining these in advance lets you set key performance indicators (KPIs), manage event tracking requirements, and measure the impact of recommendation iterations.
Example: e-commerce homepage
| Item | Value |
|---|---|
| Placement | Homepage feed |
| Recommended content | Moderated items from inventory |
| Key metrics | Add-to-cart rate, transaction volume from homepage recommendations |
| Auxiliary metrics | Click-through rate (CTR), stay duration on the recommendation page |
For this example, an add-to-cart or transaction is attributed to homepage recommendations when:
A user views the homepage feed and adds a recommended item to the cart for the first time
A user places an order from the homepage feed
A user orders an item that was recommended on the homepage, from the cart page
Step 3: Activate required cloud services
The following cloud services are required for offline modeling. Confirm that each service is activated before integration.
| No. | Cloud service | Purpose |
|---|---|---|
| 1 | MaxCompute | Data cleansing, feature engineering, and preparing training samples |
| 2 | DataWorks | Data cleansing, feature engineering, model training and evaluation, model updating, and data synchronization to online stores |
| 3 | Platform for AI (PAI) | Modeling, code editing, and scheduling of feature engineering, samples, and model training |
| 4 | Object Storage Service (OSS) | Storing checkpoint and SavedModel files of models and configuration files |
| 5 | PAI-Rec | Data diagnostics, recommendation algorithm customization, recommendation engine management, A/B testing, and report management |
After activating these services, contact an Alibaba Cloud architect to integrate them with your existing infrastructure.