models/official/recommendation/uplift

Uplift Modeling Library

Uplift modeling is a predictive modeling technique that models the incremental impact of a treatment on an individual. This repository contains a suite of tools to build, train, evaluate and export an uplift model in TensorFlow 2. All components are built in a modular fashion and integrate well with other TF-based systems such as TFX, Tensorflow Transform (TFT) and Tensorflow Model Analysis (TFMA). We use Keras as our modeling framework.

The library is divided into the following key directories:

• layers: Keras layers related to uplift modeling.
• models: Keras models that contain all the necessary logic to train, evaluate and do inference on an uplift model.
• metrics: Keras metrics used in uplift modeling.
• losses: Keras losses used in uplift modeling.

See the uplift_modeling_intro.ipynb for an introduction to uplift modeling and how to use the library to build, train and evaluate an uplift model on large scale data.

Two Tower Uplift Model

The initial release focuses on the family of models that follow a two tower uplift network architecture. The architecture draws inspiration from several related works in the field of treatment effect estimation ^{1 2 3 4 5}.

• Inputs: a mapping from feature names to feature tensors. The tensors can be of different types, eg tf.Tensor, tf.SparseTensor and tf.RaggedTensor.
• Backbone: a trainable network that computes an embedding from the inputs shared between the control and treatment arms.
• Control & Treatment Feature Encoders: trainable networks that compute embeddings from control and treatment speficic features.
• Control & Treatment Feature Combiners: methods to combine the backbone's shared embedding with the control/treatment specific embeddings.
• Control Tower: trainable network with zero or more hidden layers that learns from control examples only.
• Treatment Tower: trainable network with zero or more hidden layers that learns from treatment examples only.
• Logits Head: computes control and treatment logits. At training time, the gradient flows from the control logits for control examples and from the treatment logits for the treatment examples.

Example usage:

# Create a two tower uplift model.
uplift_network = two_tower_uplift_network.TwoTowerUpliftNetwork(
    backbone=encoders.concat_features.ConcatFeatures(
        feature_names=["shared_feature_1", "shared_feature_2"]
    ),
    treatment_feature_encoder=encoders.concat_features.ConcatFeatures(
        feature_names=["treatment_feature_1", "treatment_feature_2"]
    ),
    treatment_input_combiner=tf.keras.layers.Concatenate(),
    treatment_tower=tf.keras.Sequential([
        tf.keras.layers.Dense(64, activation="relu"),
        tf.keras.layers.Dropout(0.1),
    ]),
    control_tower=tf.keras.Sequential([
        tf.keras.layers.Dense(128, activation="relu"),
        tf.keras.layers.Dropout(0.1),
        tf.keras.layers.Dense(32, activation="relu")
    ]),
    logits_head=two_tower_logits_head.TwoTowerLogitsHead(
        control_head=tf.keras.layers.Dense(1),
        treatment_head=tf.keras.layers.Dense(1),
    ),
)
model = two_tower_uplift_model.TwoTowerUpliftModel(
  treatment_indicator_feature_name="is_treatment",
  uplift_network=uplift_network,
)

# Compile and train the model.
model.compile(
  optimizer=tf.keras.optimizers.Adagrad(learning_rate=0.05),
  loss=true_logits_loss.TrueLogitsLoss(tf.keras.losses.mean_squared_error),
  metrics=[
    treatment_fraction.TreatmentFraction(),
    uplift_mean.UpliftMean(),
    label_mean.LabelMean(),
    label_variance.LabelVariance(),
  ]
)
model.fit(dataset, epochs=10)

---

1. Johansson, Fredrik, Uri Shalit, and David Sontag. "Learning representations for counterfactual inference." International conference on machine learning. PMLR, 2016. ↩︎

2. Shalit, Uri, Fredrik D. Johansson, and David Sontag. "Estimating individual treatment effect: generalization bounds and algorithms." International conference on machine learning. PMLR, 2017. ↩︎

3. Johansson, Fredrik D., et al. "Learning weighted representations for generalization across designs." arXiv preprint arXiv:1802.08598 (2018). ↩︎

4. Hassanpour, Negar and Russell Greiner. “CounterFactual Regression with Importance Sampling Weights.” International Joint Conference on Artificial Intelligence (2019). ↩︎

5. Hassanpour, Negar and Russell Greiner. “Learning Disentangled Representations for CounterFactual Regression.” International Conference on Learning Representations (2020). ↩︎