Optimize IG Shopping ranking with multiple objectives
Overview
This question evaluates multi-objective ranking and recommender-system competencies, including predictive modeling and calibration, constrained objective aggregation, seller-level fairness constraints, fraud robustness, selection-bias-aware training, and latency-aware on-device deployment.
Meta
Oct 13, 2025, 9:49 PM
Data Scientist
Onsite
Machine Learning
0
0
Instagram Shopping: Multi-Objective Ranking With Fairness, Fraud Robustness, and On-Device Constraints
You are designing the Instagram Shopping home feed ranking to jointly optimize buyer GMV, buyer satisfaction, and seller fairness. Propose a production-ready multi-objective ranking framework that addresses:
- Modeling: Predict (a) purchase probability, (b) return risk, and (c) post-purchase satisfaction.
- Fairness: Enforce seller-level minimum exposure constraints and support cold-start seller exploration.
- Integrity: Be robust to fraudulent sellers and fake engagement.
- Latency: Support on-device prefetching with a 20 ms budget.
Describe:
- Model family and calibration (e.g., multi-task heads, calibrated scores).
- Objective aggregation (e.g., constrained optimization vs. learned utility) and how it is computed online.
- Training data sampling to counter selection bias from historical rankings.
- Post-launch monitoring for fairness and fraud drift, and how to throttle supply during spikes (e.g., flash sales).
Solution
Login required
Show