Configure all-scenario A/B experiments - Platform For AI

Global experiments allow you to configure unified parameters in the preset global scenario pairec_abtest_global_scene. These parameters are then automatically applied to all business scenarios.

All-scenario experiments

Global experiments allow you to share parameter settings across multiple recommendation scenarios and avoid repetitive configuration. PAI-Rec provides a global experiment feature to configure and apply these shared parameters from a central location. The system automatically creates a global scenario named pairec_abtest_global_scene. When you create experiments in this scenario, the configured parameters are automatically applied to all business scenarios.

How to use

In the list of scenarios, find the preset system scenario pairec_abtest_global_scene.
In this scenario, follow the standard experiment flow to create, configure, and publish an experiment. For example, you can set the parameters as follows:
```
{
    "group": 1,
    "g_weight": 0.6
}
```
After you publish the experiment, all business scenarios automatically load the experiment and its parameters when processing requests.

Note: Global experiments support multilayer, orthogonal, or mutually exclusive configurations across multiple labs. The capabilities are the same as those of standard scenarios.

Parameter rules

Unique parameter names: If a parameter is defined in a global experiment, do not configure a parameter with the same name in any business scenario experiment. If a conflict is detected during publishing, the system rejects the action and displays a notification.
No preset whitelist: You can define any parameter name. If a parameter has no corresponding logic in a specific business scenario, that scenario automatically ignores it without affecting execution.
Parameter reference support:
Business scenarios can reference global parameters in the following locations:
- Conditional expressions for traffic splitting: For example, group == 1, where group is from a global experiment.
- Experiment configuration JSON: For example, "RankScore": "(${is_click} + ${is_praise} ) * ${g_weight}", where g_weight is from a global experiment.

Activation and verification

Global experiment parameters and business scenario parameters do not overwrite each other. The final parameter set is a merge of the global and business scenario parameters.
The experiment ID for a hit request includes global experiment information, indicated by a G prefix in the ID path. For example:
..._GL101#EG11#E1002
Use the full experiment ID in logs or debugging tools to confirm that the global experiment is active.
No impact on existing experiments: The behavior of business scenarios that do not reference global parameters remains unchanged. Reporting for historical experiment data is also unaffected.

Scenarios

This feature is suitable for core policy parameters that must be unified across scenarios, such as price weights or diversity coefficients.