In the Configure Sorting Method step, find Fine Ranking and click Add in the Actions column to add a fine ranking model.
Fine Ranking Model Name: A name can contain letters and underscores (_). Recommended format: ${Scenario name}_${Model name}_rank.
Fine Ranking Target Settings: You can set multiple fine ranking destinations and destination types include classification and regression.
Classification
Fine Ranking Target Name: Specify a custom name.
Fine Ranking Target Expression: In most cases, a binary classification destination is required and aggregation is performed based on aggregation conditions.
Target Type: Set it to Classification.
Regression
Fine Ranking Target Name: Specify a custom name.
Fine Ranking Target Expression: For a regression destination, sum the event values in the behavior log table and then take the logarithm.
Target Type: Set it to Regression.
Fine Ranking Target Dependency: If a destination (x) depends on another destination (y), enter it in the Fine Ranking Target Dependency field. For example, a video is played only after a user clicks it. If the playback duration is play_time, the play_time destination can depend on the click destination. In this example, the ranking destination of the classification type is_click that is registered in the previous step is used. Other destinations of the classification type can also be used. The following figure shows the configuration.
Field to Exclude: Select features that do not need to be trained or that cannot be obtained before exposure. For example, play_time is the playback duration after exposure. The play_time, is_click, and is_buy features can lead to feature leakage if they are trained. Therefore, these features must be excluded.
Number of Training Days: indicates the number of days over which datasets are collected to train the fine ranking model.
Model Type: For single-task ranking, select a single-task model in EasyRec. For multi-task ranking, select a multi-task model. In this example, multi-task ranking is required. Select dbmtl.
Feature Selection Mode and Target Column for Feature Selection: If you set the Feature Selection Mode parameter and the Target Column for Feature Selection parameter, key features are filtered from all features. The parameters increase the model complexity. Therefore, we recommend that you do not configure the parameters in the first version of the model.
Incremental Training: indicates whether training on the current day is performed based on the model trained on the previous day. The value true indicates that incremental training is enabled. The value false indicates that full training is enabled. Default value: false.
Asynchronous Training: indicates whether asynchronous distributed training is supported.
Sample Weight: If you enable this feature, the weights of various samples are obtained based on expressions. This affects the model precision. We recommend that you do not enable this feature.
Scenarios for Sample Filtering: indicates whether data in specific scenarios is used to train models.
Feature Platform: indicates whether the configured FeatureStore is used. If you select true for this parameter and FeatureStore is configured in the environment settings of the recommendation solution, FeatureStore is used.
Automatic Feature Engineering: The value true indicates that new features are mined from existing features based on certain algorithms to facilitate model training. We recommend that you do not enable this feature in the first version of the model and enable this feature when you need to fine-tune the model.