TikTok·Data Scientist·Technical Phone Screen·Senior

Senior

May 2026

Summary

TikTok data scientist interview with a single very deep ML question about fraud detection. The whole thing was basically one giant multi-part problem and they expected you to hold all the pieces together at once, which was rough.

Questions Asked (1)

You have 10 million time-ordered transactions with 300 features and 0.5% fraud rate. Walk through your full modeling approach: which algorithm you'd start with and why (covering bias-variance tradeoffs, margins, and sensitivity to label noise on the minority class), how you'd handle class imbalance and which evaluation metric you'd prioritize, what your initial hyperparameter grid looks like with concrete starting values, how you'd design validation to avoid data leakage given the time ordering, and two scenarios where gradient boosting would actually underperform a bagged ensemble on this specific task.

Technical Trade-offsData ModelingSystem Design

Author's notes

This is basically five questions wrapped in one and they absolutely expect you to treat it that way.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by framing the problem as a high-dimensional, imbalanced, time-ordered classification task, then systematically address algorithm choice, imbalance handling, evaluation, hyperparameter tuning, and validation. Emphasize practical trade-offs and justify each decision with concrete reasoning, especially around label noise and time leakage. Conclude by contrasting gradient boosting with bagged ensembles in two specific scenarios.

Pro tip: Anchor your answer in the business context: TikTok cares about catching fraud early without disrupting user experience, so prioritize recall at a fixed precision and discuss how your choices impact real-time inference latency.

1. Algorithm Selection & Bias-Variance

Choose a starting algorithm (e.g., LightGBM or XGBoost) and justify it by discussing bias-variance trade-offs, margin maximization, and sensitivity to label noise on the 0.5% fraud class.

2. Class Imbalance & Evaluation

Explain how to handle imbalance (e.g., class weights, focal loss, resampling) and select an appropriate metric (e.g., PR-AUC, recall at high precision) given the fraud detection goal.

3. Hyperparameter Grid

Propose a concrete initial hyperparameter grid for the chosen algorithm, with starting values for learning rate, tree depth, regularization, and subsampling.

4. Time-Aware Validation

Design a validation scheme that respects time ordering, such as expanding window or rolling-origin cross-validation, to prevent data leakage.

5. When Boosting Underperforms Bagging

Describe two scenarios where gradient boosting would underperform a bagged ensemble (e.g., random forest) on this specific task, such as high label noise or need for stable, low-variance predictions.

Key Points to Mention

Bias-variance trade-off: boosting reduces bias but can overfit noise; bagging reduces variance.
Margin maximization: boosting focuses on hard examples, which may be mislabeled in fraud detection.
Label noise sensitivity: boosting can overfit to noisy minority class labels; consider robust loss functions or noise-robust algorithms.
Class imbalance: use class weights, focal loss, or stratified sampling; avoid naive oversampling that duplicates noise.
Evaluation metric: prioritize PR-AUC or recall at 90% precision over ROC-AUC due to imbalance.
Time-based validation: use expanding window or rolling-origin CV to mimic real-world deployment and avoid leakage.

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