Skip to main content

Wide & Deep (Neural Scoring)

Retired

This policy is retired and is no longer accepted in setup_engine or update_engine requests via the V2 API. Existing engines using this policy will continue to serve inference. Please reach out if you'd like to use this policy and we can add it back.

Description​

The Wide & Deep policy implements the Wide & Deep Learning model, which jointly trains a wide linear model and a deep neural network (DNN) to combine the benefits of memorization and generalization.

  • Wide Part: A generalized linear model often fed with raw sparse features and manually engineered cross-product features to effectively memorize specific feature combinations.
  • Deep Part: A standard MLP fed with dense embeddings (learned from categorical features) and potentially normalized dense features to generalize by learning complex, non-linear patterns.

The outputs are typically summed before the final prediction layer.

Policy Type: wide-deep Supports: scoring_policy

Premium Model

This model requires the Standard Plan or higher.

Hyperparameter tuning​

  • val_split: Fraction of training data to use for validation.
  • n_epochs: Number of complete passes through the training data.
  • num_workers

V1 API

policy_configs:
  scoring_policy:
    policy_type: wide-deep
    # Architecture
    deep_hidden_units: [256, 128, 64] # Layer sizes for the deep MLP component
    activation_fn: "relu"             # Activation for deep layers (e.g., "relu", "sigmoid")
    # Training Control
    val_split: 0.1                    # Proportion of data for validation during training
    n_epochs: 10                      # Number of training epochs

Usage​

Use this model when:​

  • You want to combine memorization of specific feature combinations with generalization to unseen patterns
  • You have the ability to engineer meaningful cross-product features for the wide part
  • You have rich categorical features that can be embedded for the deep part
  • You are building recommendation or search ranking systems similar to app store or content feeds

Choose a different model when:​

  • You do not want to maintain manual feature engineering for the wide component (use DeepFM instead)
  • You prefer tree-based models with simpler feature engineering workflows (use LightGBM or XGBoost)
  • You primarily need retrieval-stage embeddings rather than point-wise scoring (use Two-Tower, ALS, or BeeFormer)
  • You need to model detailed sequential user behavior (use SASRec, BERT4Rec, or other sequential models)

Use cases​

  • App recommendation and store ranking (apps, games, extensions)
  • Personalized content ranking for feeds, homepages, and carousels
  • Search ranking where memorized query–item patterns and generalization both matter
  • CTR prediction and conversion modeling in environments with rich, hand-crafted cross-features

Reference​