Skip to main content

Hyperparameter tuning

Shaped automatically tunes model hyperparameters to optimize performance. You can configure hyperparameter tuning by specifying exact values, setting min/max ranges for partial tuning, or letting Shaped handle all tuning automatically.

Autotuning by default​

If you don't specify hyperparameters in your model configuration, Shaped automatically tunes all tunable hyperparameters. This applies to all model policies.

training:
  models:
    - name: my_model
      policy_type: lightgbm
      # No hyperparameters specified - all will be autotuned

Event values filtering​

At minimum, you can specify event_values to filter which interaction events are used for training. This is useful when your interaction table contains multiple event types and you want to train on specific events.

training:
  models:
    - name: purchase_model
      policy_type: lightgbm
      event_values:
        - purchase
        - checkout_complete

Partial tuning​

Many hyperparameters support partial tuning by specifying min and max values. Shaped will tune the parameter within the specified range.

LightGBM example​

training:
  models:
    - name: lightgbm_model
      policy_type: lightgbm
      max_depth:
        type: tunable_int
        min: -1
        max: 10
      num_leaves:
        type: tunable_int
        min: 20
        max: 40
      learning_rate:
        type: tunable_float
        min: 0.001
        max: 0.1

ELSA example​

training:
  models:
    - name: elsa_model
      policy_type: elsa
      factors:
        type: tunable_int
        min: 10
        max: 200
      lr:
        type: tunable_float
        min: 0.01
        max: 0.1

BERT4Rec example​

training:
  models:
    - name: bert4rec_model
      policy_type: bert4rec
      batch_size:
        type: tunable_int
        min: 32
        max: 2048
      learning_rate:
        type: tunable_float
        min: 0.0001
        max: 0.1
      attn_dropout_prob:
        type: tunable_float
        min: 0.0
        max: 0.5
      hidden_dropout_prob:
        type: tunable_float
        min: 0.0
        max: 0.5

Combining approaches​

You can combine exact values, partial tuning, and autotuning in the same model configuration. Any hyperparameters not specified will be autotuned.

training:
  models:
    - name: tuned_model
      policy_type: lightgbm
      event_values:
        - click
        - view
      max_depth:
        type: tunable_int
        min: 5
        max: 8
      # learning_rate, num_leaves, and other hyperparameters will be autotuned

End-to-end example​

This example shows how to train an ELSA and LightGBM model with hyperparameter tuning, then combine their scores in a query.

Engine configuration​

data:
  item_table:
    name: products
    type: table
  user_table:
    name: users
    type: table
  interaction_table:
    name: interactions
    type: table

training:
  models:
    - name: elsa_model
      policy_type: elsa
      factors:
        type: tunable_int
        min: 10
        max: 200
      lr:
        type: tunable_float
        min: 0.01
        max: 0.1
    - name: lightgbm_model
      policy_type: lightgbm
      event_values:
        - click
        - purchase
      max_depth:
        type: tunable_int
        min: 5
        max: 10
      num_leaves:
        type: tunable_int
        min: 20
        max: 40
      learning_rate:
        type: tunable_float
        min: 0.001
        max: 0.1

Query with combined scoring​

Use a value model expression in the score stage to combine outputs from both models:

queries:
  personalized_feed:
    query:
      type: rank
      from: item
      retrieve:
        - type: column_order
          columns:
            - name: _derived_popular_rank
              ascending: true
          limit: 1000
      score:
        type: score_ensemble
        value_model: 0.6 * elsa_model + 0.4 * lightgbm_model
        input_user_id: $parameters.user_id
      limit: 20
    parameters:
      user_id:
        default: null

The value model expression 0.6 * elsa_model + 0.4 * lightgbm_model weights the ELSA model at 60% and the LightGBM model at 40%. Adjust these weights based on your performance requirements.