Bytedance·Machine Learning Engineer·Technical Phone Screen·Intermediate

Intermediate

Apr 2026

Summary

Bytedance ML engineer interview that went pretty deep into fundamentals. They wanted a full logistic regression implementation from scratch, not just theory, so if you're prepping for this role make sure you can actually write the math out in code.

Questions Asked (1)

Implement logistic regression from scratch: write the sigmoid function, binary cross-entropy loss, gradients for weights and bias, a training loop with configurable learning rate and iterations, plus predict_proba and predict methods. Optionally add L2 regularization and explain what it does.

Algorithms & Data StructuresTechnical Trade-offs

Author's notes

This was more involved than I expected for a phone screen.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by writing the core functions (sigmoid, loss, gradients) with clear vectorized implementations, then build the training loop with configurable hyperparameters, and finally add predict methods. Optionally implement L2 regularization and explain its effect on the loss and gradients.

Pro tip: Emphasize vectorization for efficiency and numerical stability (e.g., clipping sigmoid inputs) to show production-level awareness. Also, mention that L2 regularization is equivalent to adding a Gaussian prior on weights, which helps prevent overfitting.

1. Implement core functions

Write the sigmoid function and binary cross-entropy loss, ensuring numerical stability by avoiding overflow in exponentials.

2. Compute gradients

Derive and implement the gradients of the loss with respect to weights and bias, using vectorized operations for efficiency.

3. Build training loop

Create a training loop that iteratively updates weights and bias using gradient descent, with configurable learning rate and number of iterations.

4. Add prediction methods

Implement predict_proba to return probabilities and predict to return binary class labels based on a threshold (e.g., 0.5).

5. Optional: L2 regularization

Add L2 regularization to the loss and gradients, and explain that it penalizes large weights to reduce overfitting.

Key Points to Mention

Vectorization for efficient computation
Numerical stability in sigmoid and loss (e.g., log-sum-exp trick)
Gradient derivation and vectorized implementation
Configurable hyperparameters (learning rate, iterations)
L2 regularization: adds penalty term to loss and gradient, controls overfitting
Threshold choice for predict method and its impact

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