Ranking

Onboarding a Ranking Model

Suppose you would like to onboard a ranking model for the following dataset.

Following is an example of how you would construct a fdl.ModelInfo object for a ranking model.

PROJECT_ID = 'example_project'
DATASET_ID = 'expedia_data'
MODEL_ID = 'ranking_model'

model_task = fdl.ModelTask.RANKING
model_group_by = 'srch_id'
model_target = 'click_bool'
model_outputs = ['score']
raning_top_k = 20
model_features = [
    'price_usd',
    'promotion_flag',
    'weekday',
    'week_of_year',
    'hour_time',
    'minute_time'

model_info = fdl.ModelInfo.from_dataset_info(
    dataset_info=dataset_info,
    dataset_id=DATASET_ID,
    features=model_features,
    group_by=model_group_by,
    ranking_top_k=ranking_top_k,
    target=model_target,
    outputs=model_outputs,
    model_task=model_task,
)

client.add_model(
    project_id=PROJECT_ID,
    dataset_id=DATASET_ID,
    model_id=MODEL_ID,
    model_info=model_info
)

📘

Note

Using client.add_model() does not provide Fiddler with a model artifact. Onboarding a model in this fashion is a good start for model monitoring, but Fiddler will not be able to offer model explainability features without a model artifact. You can subsequently call client.add_model_surrogate() or client.add_model_artifact() to provide Fiddler with a model artifact. Please see Uploading a Model Artifact for more information.

🚧

Note

group_by: when onboarding a ranking model, you must specify a group_by argument to the fdl.ModelInfo object. It will tell Fiddler which column should be used for grouping items so that they may be ranked within a group.

ranking_top_k: an optional parameter unique to ranking model. Default to 50. It's an int representing the top k outputs to take into consideration when computing performance metrics MAP and NDCG.

📘

Tips

When onboarding a graded ranking model with categorical target, categorical_target_class_detail is a required argument for fdl.ModelInfo object. For example: categorical_target_class_details=['booked','click_no_booking','no_click']