12 TikTok Machine Learning Engineer (New Grad) Interview Questions (2026)

Outline

Tie to wanting production ML impact — models that affect billions of users daily, tight loop between research and deployment, deep feature-engineering opportunities. TikTok's recommendation system is one of the most studied in the world; that's a draw. Avoid 'ML is interesting' — too generic.

Outline

STAR. Pick a real story. Common causes: train/serve skew (different feature computation), data leakage in training, label distribution shift, sampling bias in eval set. Show your diagnostic process — comparing training and serving features, A/B testing, error analysis.

Outline

Pick the project you can defend in extreme detail. Be ready for: data, architecture, training loop, hyperparameters, eval metric, failure modes, what you'd do differently. TikTok MLE interviewers drill — surface project chat fails. Recommendation, ranking, or retrieval projects fit best; CV/NLP/RL projects also work.

Outline

Quickselect — partition around random pivot, recurse on relevant side. Average O(N), worst O(N²). Alternative: min-heap of size k, O(N log k). TikTok expects quickselect for the bonus.

Outline

cos_sim = (a . b) / (||a|| * ||b||). One pass to compute dot product and both magnitudes. Handle zero-vector edge case (return 0 or NaN, depending on convention). O(N) time, O(1) space. TikTok's recommendation system uses cosine similarity for embedding retrieval.

Outline

Hash user IDs into experiment buckets. Treatment gets new model; control gets baseline. Measure primary metric (e.g., daily watch time per user) and guardrails (latency, error rate, user-complaint rate). Run for enough traffic-days to reach significance. Discuss novelty effect (treatment looks good first week then decays) and interaction effects with concurrent experiments.

Outline

Multi-objective: predict P(watch), P(like), P(share), P(complete), P(comment). Combine into a single utility score with weights. Model: deep neural network with embeddings for user, video, creator, and contextual features. Training: log past interactions, label is the observed action. Discuss feature freshness, online vs offline training, serving latency budget (10s of ms for ranking). TikTok's domain — go deep.

Outline

Collaborative filtering = recommend based on other users' behavior (users who liked X also liked Y). Cold-start hard for new users/items. Content-based = recommend based on item features (videos similar to ones you watched). Cold-start easier. Modern systems use hybrid — embeddings learned from both behavior signals and content features.

Outline

Logistic regression = linear model with sigmoid output, no hidden layer, single decision boundary. NN with one hidden layer = nonlinear basis functions (hidden units) before final output — can fit nonlinear boundaries. Universal approximator with enough hidden units. Trade off: more parameters, harder to interpret, easier to overfit.

Outline

Detect first: monitor feature distributions in production, alert on drift (KL divergence, population stability index). Causes: external event, upstream pipeline change, adversarial behavior. Mitigations: retrain on recent data, rollback to a stable model, hot-fix bad features. Discuss the tradeoff between fast response and overfitting to noise.

Outline

Two encoders (user tower, item tower) produce embeddings in shared space. Trained so positive pairs (user, item user interacted with) have high dot product. At serve time: compute user embedding online (fast), look up nearest item embeddings via approximate nearest neighbor (FAISS, ScaNN). Decouples the heavy item encoding (precomputed) from per-request user encoding. Standard for retrieval at TikTok-scale.

Outline

Random negatives (uniform from item catalog): simple, weak signal. In-batch negatives: other users' positives in the same training batch — efficient, popular. Hard negatives: items similar to positive but not interacted (mined from current model embeddings). Mix all three. TikTok's recall stage relies on this.