Skip to main content

Score Ensemble (Special)

Description​

The Score Ensemble policy combines the ranked lists generated by multiple underlying scoring policies. The primary mechanism is typically interleaving, taking top items from each constituent policy's ranked list to create the final combined list. This leverages the diverse strengths of different models, potentially improving overall ranking quality, robustness, or diversity.

Policy Type: score-ensemble Supports: scoring_policy

Configuration Example​

scoring_policy_ensemble.yaml
policy_configs:
  scoring_policy:
    policy_type: score-ensemble
    # List the full configurations of the policies to include in the ensemble
    policies:
       - policy_type: lightgbm # Policy 1 configuration
         objective: lambdarank
         n_estimators: 500
         # ... other params
       - policy_type: sasrec   # Policy 2 configuration
         hidden_size: 128
         n_layers: 2
         # ... other params
       - policy_type: popular # Policy 3 configuration
         mode: sum

References​

  • Dietterich, T. G. (2000). Ensemble methods in machine learning. Multiple Classifier Systems.
  • Macdonald, C., et al. (2012). Assessing the utility of score distributions for rank fusion. CIKM. (Discusses rank fusion/interleaving).