UpliftRecommender¶

The UpliftRecommender is designed for causal inference scenarios where you want to estimate the incremental impact (uplift) of recommending an item, rather than just predicting overall engagement.

Overview¶

Purpose: Maximize the causal effect of recommendations by estimating how much more likely a user is to engage because of the recommendation, not just correlation.

Key Insight: Traditional recommenders might recommend items users would engage with anyway. Uplift modeling focuses on items where the recommendation makes the biggest difference.

When to Use¶

✅ Perfect For: - Maximizing incremental revenue (not total revenue) - Understanding true treatment effects of recommendations - Avoiding "preaching to the choir" (recommending what users would choose anyway) - Marketing campaigns where you want to maximize lift - Resource-constrained scenarios (limited promotion slots)

❌ Not Ideal For: - Standard ranking tasks (use RankingRecommender) - When you don't have treatment/control data

Causal Inference Strategies¶

1. T-Learner (Two-Model Approach)¶

Trains separate models for treatment and control groups
Uplift = Prediction(treatment) - Prediction(control)

2. S-Learner (Single-Model Approach)¶

Trains one model with treatment indicator as a feature
Uplift = Prediction(treated=1) - Prediction(treated=0)

Basic Usage¶

from skrec.estimator.classification.xgb_classifier import XGBClassifierEstimator
from skrec.recommender.uplift_model.uplift_t_learner import UpliftTLearnerRecommender
from skrec.scorer.universal import UniversalScorer

# Create T-Learner uplift recommender
estimator = XGBClassifierEstimator({"learning_rate": 0.1, "n_estimators": 100})
scorer = UniversalScorer(estimator)
recommender = UpliftTLearnerRecommender(scorer)

# Train (requires treatment indicator in interactions data)
recommender.train(
    interactions_ds=interactions_with_treatment_ds,
    users_ds=users_dataset,
    items_ds=items_dataset
)

# Recommend based on uplift scores
recommendations = recommender.recommend(
    interactions=interactions_df,
    users=users_df,
    top_k=5
)

Data Requirements¶

Your interactions dataset must include a treatment indicator:

interactions_df = pd.DataFrame({
    "USER_ID": ["user_1", "user_2", "user_3"],
    "ITEM_ID": ["item_A", "item_B", "item_C"],
    "OUTCOME": [1.0, 0.0, 1.0],
    "treatment": [1, 0, 1]  # 1=treated (saw recommendation), 0=control
})

Example: Incremental Revenue¶

# Traditional recommender might recommend item_A
# because it has high overall conversion rate (80%)
# 
# But what if:
# - 75% of users buy item_A WITHOUT recommendation
# - Only 80% buy it WITH recommendation
# - Uplift = 80% - 75% = 5% (small incremental impact)
#
# Meanwhile item_B:
# - 10% of users buy item_B WITHOUT recommendation
# - 40% buy it WITH recommendation  
# - Uplift = 40% - 10% = 30% (large incremental impact!)
#
# Uplift recommender will prefer item_B!

Best Practices¶

1. Data Collection¶

Collect randomized treatment/control data
Ensure sufficient samples in both groups
Avoid selection bias

2. Model Selection¶

T-Learner: Better when treatment effect is heterogeneous
S-Learner: Simpler, works well with abundant data

3. Evaluation¶

Measure incremental lift, not absolute performance
Use A/B testing to validate uplift estimates
Compare against baseline (no treatment)

4. Interpretation¶

Negative uplift = Item hurts when recommended (avoid!)
Zero uplift = Recommendation has no effect
Positive uplift = Recommendation helps (prioritize!)

Common Use Cases¶

Marketing Campaigns¶

# Recommend items with highest email campaign lift
# Don't waste email slots on items users would buy anyway
uplift_recommendations = recommender.recommend(...)

Promotional Budgets¶

# Allocate discounts to items with highest uplift from promotion
# Save money by not discounting items that would sell anyway

Treatment Effect Heterogeneity¶

# Estimate uplift varies by user segment
# Young users might respond differently to recommendations than older users

Implementation Details¶

For implementation details and API reference, see: - skrec/recommender/uplift_model/uplift_recommender.py - skrec/recommender/uplift_model/uplift_scorer_adapter.py