ByteDance Data Scientist Interview Questions

ByteDance·Data Scientist·Technical Phone Screen

Jul 2026

Bytedance data scientist interview that leaned heavily on stats fundamentals. The whole session felt like a hypothesis testing deep dive, which I wasn't fully expecting going in.

Define Type I and Type II errors in the context of binary classification and hypothesis testing.
How do Type I and Type II errors map onto false positives, false negatives, precision, and recall?
When is a Type I error more costly than a Type II error, and when is the reverse true? Give concrete examples.
Given a specific scenario like fraud detection or medical screening, identify which error type corresponds to which mistake.

“Started fine, rattled off the textbook definitions.”

View Post

ByteDance·Data Scientist·Technical Phone Screen

Jul 2026

Bytedance Data Scientist interview with a coding problem on parentheses balancing. Pretty standard algorithmic screen but the problem had a small wrinkle that tripped me up at first.

Given a string of only '(' and ')' characters, what is the minimum number of parentheses you need to insert to make the string valid? Walk through your approach and its time and space complexity.

“My first instinct was to just count unmatched brackets with a stack, which got me partway there.”

View Post

ByteDance·Data Scientist·Technical Phone Screen

Jul 2026

Interviewed for a Data Scientist intern role at Bytedance on their short-video platform side. Three meaty prompts back to back, all centered on a real project from my resume. The product research and creator engagement questions were the ones that tripped me up the most.

Walk us through a recent data or ML project: the business problem, the data, your role, how you handled preprocessing and the bias-variance tradeoff, and how you chose your evaluation metrics and tied them back to the business goal.
If the company were to launch the feature or product you built, how would you design the research plan before and after launch? Cover your hypotheses, user segments, data needs, experiment design, success metrics, guardrail metrics, and how you'd make the go/no-go call.
You have a metric tracking new users who registered and posted at least once on the same day. The company wants to improve creator engagement. How do you analyze this and what do you recommend?

“I picked a recommendation model project and it went okay until they pushed on the bias-variance piece.”

View Post

ByteDance·Data Scientist·Technical Phone Screen

Jun 2026

Bytedance Data Scientist interview with a SQL question that looked straightforward but had a subtle deduplication requirement baked in. The kind of problem where you can get a working query pretty fast and still miss the point entirely.

Given a users table with registration timestamps and a posts table with post creation timestamps, write a SQL query that returns, for each calendar date, the count of distinct users who both registered on that date and published at least one post on that same date.

“The dedup part is where people probably slip up.”

View Post

ByteDance·Data Scientist·Technical Phone Screen

Jun 2026

Bytedance data scientist interview with a meaty ML question on classification thresholds. One question but it had a bunch of sub-parts that each could've been their own conversation.

You've built a binary classifier that outputs probabilities. Walk through how you'd pick a decision threshold on a validation set, how you'd factor in asymmetric costs between false positives and false negatives, how class imbalance and calibration issues affect your choice, and give examples of when you'd want a high vs low threshold.

“This question looks like one thing but it's really four.”

View Post

ByteDance·Data Scientist·Technical Phone Screen

Jun 2026

SQL round for a Data Scientist role at Bytedance. One question, medium difficulty, and the interviewer nudged me about duplicates partway through which I probably should've caught myself.

Write a SQL query to find the total number of distinct users who both registered and published at least one post on the same day. Keep in mind a user can publish multiple posts in a single day.

“The interviewer had to remind me about duplicates, which was a bit embarrassing.”

View Post

ByteDance·Data Scientist·Technical Phone Screen

Jun 2026

New-grad ML/data scientist loop at ByteDance. Six questions back to back, all technical, no fluff. The range was pretty wide: systems, retrieval, classic ML theory, and LLM alignment stuff. Left feeling like I'd studied the wrong half of the syllabus.

You need to deploy a multimodal model (text plus image, video, or audio) under tight GPU memory, latency, and cost constraints. How would you redesign both the model and the serving infrastructure to hit those constraints without wrecking quality?
You have a large video corpus where each video already has a caption and precomputed embeddings. How would you build a retrieval system that returns relevant results fast and with high recall?
What is overfitting, how do you detect it, and what are the most effective ways to reduce it in deep learning?
Explain the intuition behind dropout and why the inverted dropout trick keeps activation magnitudes consistent between training and inference. When does dropout hurt more than it helps?
Compare Batch Normalization, Layer Normalization, Group Normalization, Instance Normalization, and RMSNorm. What statistics does each use, how do they behave differently at training vs inference, and why do transformers tend to prefer certain ones?
Walk through how reinforcement learning is used to fine-tune LLMs after supervised training. Cover the reward model, KL regularization, common failure modes, and how you'd evaluate the result. How does PPO-based alignment compare to direct preference optimization approaches?

“This one sprawled in every direction and I didn't manage it well.”

View Post

ByteDance·Data Scientist·Technical Phone Screen

May 2026

Bytedance data scientist interview with a classic probability geometry problem. Pretty clean question but the derivation takes a minute to set up properly if you haven't seen it before.

A stick of length 1 is broken at two points chosen independently and uniformly at random. What is the probability that the three resulting pieces can form a triangle?

“I'd seen this problem before so I knew the answer was 1/4, but actually deriving it cleanly under pressure is another thing.”

View Post

ByteDance·Data Scientist·Technical Phone Screen

May 2026

Bytedance Data Scientist interview with a coding-heavy session that escalated from stack design to streaming statistics to tree problems. The progression felt deliberate, like each question was trying to build on the last, which was an interesting structure but also kind of exhausting.

Design a stack that supports push, pop, top, and getMin, where all operations run in O(1) time.
Design a stack that also supports peekMax (return the current max without removing it) and popMax (remove and return the maximum element), with efficient time complexity.
Design a data structure that supports adding numbers from a stream one at a time and returning the current median after each insertion. Also discuss whether the MaxStack design idea can be adapted here.
Given a binary tree with integer values, determine if there exists a path that sums to a target value, where the path can start at any node, moves only upward toward ancestors, and must be a contiguous chain of nodes (no branching, no revisiting).

“Classic warmup.”

View Post

ByteDance·Data Scientist·Online Assessment (OA)

May 2026

Bytedance Data Scientist online assessment with three coding problems. Nothing behavioral, just straight algorithmic stuff. The problems ranged from trivially easy to genuinely tricky depending on how you think about segment maintenance.

Given a string, treat uppercase and lowercase as identical. Count how many adjacent pairs of characters differ from each other.
Implement a memory allocator that only starts allocations at indices divisible by 8. Support alloc(x) to find the leftmost valid free block and assign it an ID, and erase(id) to free all cells belonging to that ID.
Points on a number line can be colored one at a time. After each update (assign or change a color), return the total number of maximal contiguous same-color segments. Design an efficient online approach.

“Easiest of the three.”

View Post