Artificial Intelligence Recommendation:Sorting configuration

In this step, add and configure a fine-ranking model.

In this example, set Fine Ranking Model Name to acc_rank. The Fine Ranking Target Settings includes two targets: is_click with the expression max(if(event='click', 1, 0)), and is_praise with the expression max(if(event='praise', 1, 0)). For Fields to Exclude, select fields such as age. Set Number of Training Days to 30. For Model Type, select dbmtl, and for the embed_dim strategy, select EMB_SQRT4_STEP4.

• Fine Ranking Model Name: A string that contains only letters and underscores (_). The recommended format is ${scenario_name}_${model_name}_rank.

• Fine Ranking Target Settings: You can define one or more classification or regression targets.

Classification targets

You can specify a custom name for the fine-ranking target. For the fine-ranking target expression, you typically define a binary classification target by using an aggregate function. The target type is classification.

For example, set Fine Ranking Target Name to is_click and Fine Ranking Target Expression to max(if(event='click', 1, 0)). The Fine Ranking Target Dependency (depends) is optional. If you have multiple dependencies, separate them with commas.

Regression targets

You can specify a custom name for the fine-ranking target.

For a regression target, the expression must sum the target numeric values from the behavior log table and then apply a log transformation to the result.

For Target Type, select regression.

Use the Fine Ranking Target Dependency setting when one target (X) can occur only after another target (Y). For example, in a video scenario, playback can happen only after a click. Therefore, the play_time target can depend on the is_click target, where is_click is the name of the classification task defined previously. This principle also applies to other classification tasks.

• Fields to Exclude: Select features to exclude from training. Exclude any feature that is unavailable before an exposure or could cause feature leakage. For example, avoid including fields such as play_time (the user's watch time after an exposure) or other post-interaction data like click or payment status, as this can cause feature leakage.

• Number of Training Days: Specify the number of days of data from the dataset to use for training the fine-ranking model.

• Model Type: Select a single-task model from EasyRec if you have a single target. If you have multiple targets, select a multi-task model. In this example, because we have multiple tasks, we select dbmtl.

• Feature Selection Mode and Target Column for Feature Selection: When enabled, this option filters available features to select only the most important ones for training. The target column works with the selected mode to choose these features. This feature can complicate the initial setup, so we recommend not using it for your first model.

• Incremental Training: Set to false for full training, which is the recommended default. Setting this to true enables incremental training, where the next day's training job continues from the previous day's model checkpoint.

• Asynchronous Training: Specify whether to use asynchronous training in a distributed training environment.

• Sample Weight: When configured, this setting assigns weights to samples based on an expression, which can affect model accuracy. This feature is generally not used.

• Scenarios for Sample Filtering: Specify whether to use data from a specific scenario for training.

• Feature Platform: Specifies whether to use a configured FeatureStore. This option is enabled if you have configured a FeatureStore in the environment settings of your recommendation solution.

• Automatic Feature Engineering: This feature uses algorithms to generate more effective features from existing ones for training. We recommend not using automatic feature engineering (autofe) for your first model. Enable it later to improve model performance.