TikTok·Data Scientist·Technical Phone Screen·Intermediate

Intermediate

Jun 2026

Summary

TikTok data scientist interview with a machine learning case question centered on churn prediction. Pretty standard stuff if you've done this type of modeling before, but there's more to unpack than it looks.

Questions Asked (1)

You need to build a churn prediction model for a subscription app. What modeling approaches would you consider, and how would you handle class imbalance and pick the right evaluation metrics?

Product Analytics & MetricsTechnical Trade-offsData Modeling

Author's notes

I started with logistic regression as a baseline because it's fast and interpretable, then moved into tree-based models like gradient boosting.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by framing churn as a binary classification problem and discuss a range of models from logistic regression to gradient boosting, emphasizing trade-offs. Then explain how you'd address class imbalance using techniques like class weighting, resampling, or anomaly detection, and finally justify evaluation metrics like PR-AUC, recall at fixed precision, or lift, aligning with business goals.

Pro tip: Tie your metric choice directly to business impact—e.g., 'If retention offers are cheap, optimize for recall; if expensive, optimize for precision.' This shows you think beyond technical metrics.

1. Clarify the problem and data

Define churn precisely (e.g., no activity for X days) and consider data available: user behavior, engagement, tenure. Mention that churn is often rare and time-dependent.

2. Choose modeling approaches

Discuss interpretable models (logistic regression) for baseline and complex models (XGBoost, LightGBM, neural networks) for performance. Consider survival analysis for time-to-churn.

3. Handle class imbalance

Explain techniques: class weights, oversampling (SMOTE), undersampling, or anomaly detection. Note that imbalance handling should be validated via cross-validation.

4. Select evaluation metrics

Move beyond accuracy: use PR-AUC, recall at high precision, F1, or lift. Align with business costs of false positives/negatives.

5. Validate and iterate

Use time-based validation to mimic production. Monitor performance and retrain periodically. Consider model explainability for stakeholder buy-in.

Key Points to Mention

Churn definition and prediction window (e.g., predict churn in next 30 days)
Model trade-offs: interpretability vs. performance, training time, scalability
Class imbalance techniques: class weights, SMOTE, undersampling, and their pros/cons
Evaluation metrics: PR-AUC, recall@precision, lift, and business-aligned metrics
Time-based validation and concept drift
Feature engineering: recency, frequency, monetary (RFM), engagement trends

AI-generated suggestions, not part of the candidate's original notes. May be inaccurate — verify before relying on them.