Design an ad-selection system across objectives
Overview
This question evaluates expertise in machine learning system design for ad selection and real-time bidding, covering predictive modeling and calibration, value unification across objectives, auction mechanics, budget pacing, exploration strategies, guardrails, and evaluation methodologies.
TikTok
Oct 13, 2025, 9:49 PM
Data Scientist
Technical Screen
Machine Learning
0
0
End-to-End Ad-Selection System Design
Context
You must choose, at impression time, which advertiser type to show to a user. There are three advertiser objectives with fixed value schemes (values reflect the platform's objective weights per action):
- Gaming (CPI): $10 per install
- E‑commerce (CPA): $5 per purchase
- Branding (CPM): $3 per 1000 impressions
Design a system that maximizes expected advertiser value per impression while respecting budgets and pacing.
Requirements
Specify the following components:
-
Prediction
- Features and models to estimate P(install | ad, user, context) and P(purchase | ·).
- Calibration/debiasing for auction-time use.
- Updating cadence.
-
Value Unification and Constraints
- How to make scores commensurate across objectives (e.g., expected value per impression; convert CPM to per‑impression value).
- Incorporate predicted win rate, quality thresholds, and user experience constraints.
-
Exploration, Cold Start, Guardrails
- Exploration vs. exploitation (e.g., contextual bandits).
- Cold‑start handling.
- Guardrails (frequency capping, diversity, fairness).
-
Auction and Budgeting
- Auction mechanics (second‑price/VCG), bid shading, pacing control.
- Multi‑objective bidding interactions.
-
Evaluation
- Offline/online evaluation: counterfactual replay, lift tests.
- Success metrics (incremental ROAS, retention, long‑term engagement).
- Detecting/mitigating feedback loops or gaming.
Solution
Login required
Show