Stageboost: Recommending Signals Based on Counterfactual Estimation
Paper Guide Brief
Reading Brief
Stageboost is a two-stage XGBoost-based model for personalized signal assignment on eBay's View-Item page. It first predicts baseline conversion probability from item/user features, then estimates signal-specific uplift using differential feature weighting and feature subsampling within a single gradient boosting ensemble. Trained on randomized controlled trial data, it uses a counterfactual offline evaluation metric and achieved a 0.08% GMB lift overall and 0.58% in a focus category in online A/B tests.
Central Claim
Introduces a novel two-stage XGBoost approach for personalized signal assignment in e-commerce, combining baseline conversion modeling with signal-specific uplift estimation in a single model, and demonstrates its effectiveness through counterfactual evaluati...
Contribution
Introduces a novel two-stage XGBoost approach for personalized signal assignment in e-commerce, combining baseline conversion modeling with signal-specific uplift estimation in a single model, and demonstrates its effectiveness through counterfactual evaluation and online experiments.
Why It Matters
The paper presents a unified two-stage gradient boosting framework with differential feature weighting and feature subsampling to jointly model baseline conversion and signal-specific uplift, addressing the limitations of deterministic assignment and data sparsity in prior work.
Prerequisites
XGBoost, two-stage modeling, uplift modeling, counterfactual estimation, feature weighting
Atlas Placement
Information Retrieval (subfield)
Read If
You care about XGBoost, two-stage modeling, uplift modeling.
Skip If
You only care about counterfactual conversion rate, Gross Merchandise Bought (GMB).
Noosaga Placements
- The paper addresses the problem of selecting and ranking signals for display on an e-commerce page, which is a form of information retrieval and ranking task.the goal is to select and rank a subset R ⊂ Q where |R| = min(k_i, |Q|) to maximize the expected conversion probabilitywe aim to learn a ranking function f : X × S → R that assigns a score to each (context, signal) pair
- Ensemble Methodsframework90%The model is based on XGBoost, which is an ensemble method using gradient boosting of decision trees.we present a 2 stage xgboost based modelStageboost is a two-stage gradient boosting approachXGBoost’s feature importance weighting mechanism
- The core methodology is a two-stage XGBoost-based supervised learning approach for uplift modeling and prediction.Stageboost is a two-stage gradient boosting approachwe train a single XGBoost classifier on the full feature setM1 (Baseline Conversion Model) learns to predict the baseline probability of conversion
- Supervised Learningframework80%The model is trained on labeled conversion data in a supervised manner to predict binary conversion outcomes.Y ∈ 0, 1 is the binary conversion indicatorwe train a single XGBoost classifier
- Causal Inferenceframework70%The paper uses counterfactual estimation and uplift modeling to estimate the causal effect of showing signals on conversion.estimates the incremental uplift in conversion probability attributable to showing each specific signalthe counterfactual conversion rate for this group is estimated using the adjustment formulaCausal estimation [5, 10] and uplift modeling [4, 6] approaches have been successfully applied
- The work applies AI techniques to a real-world e-commerce personalization problem, and discusses future work in reinforcement learning.with plans to explore RL [7] in future workFuture work includes exploring reinforcement learning approaches for online adaptation
- Statistical Learning Theoryframework50%The paper discusses statistical significance and uses randomized controlled trials to ensure unbiased training data, relating to statistical learning principles.Randomization ensures uniform exploration, providing balanced training databased on p-value
Abstract
Signals are short textual or visual snippets displayed on the eBay View-Item (VI) page, providing additional, contextual information for users about the viewed item. The aim of displaying these signals is to facilitate intelligent purchase and to incentivize engagement. In this paper, we present a 2 stage xgboost based model that optimally populates the VI page with signals. This approach has shown a 0.08% lift in overall GMB (Gross Merchandise Bought) and 0.58% increase in Parts and Accessories GMB, primarily due to increase in conversion of high average price items in online experimentation.
Paper Context
Classified from the full extracted paper text (23,476 characters). The Paper Guide brief above is the user-facing synthesis; raw context is kept out of the page.
Full-paper context sent 23,476 of 23,476 extracted characters to classification.