← Bytedance Interview Insights

Bytedance·Machine Learning Engineer·Onsite - System Design / Architecture·Senior

Senior
May 2026

Summary

Bytedance ML Engineer interview with a meaty system design question around fraud detection in e-commerce. Pretty open-ended, which I wasn't fully prepared for.

Questions Asked (1)

Q1

Design a machine learning system to detect self-dealing or wash transactions in an e-commerce marketplace, where sellers use related buyer accounts to fake sales volume, inflate rankings, or generate fraudulent reviews.

System DesignData ModelingTechnical Trade-offs
Author's notes

This one sprawled in every direction.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by clarifying the business objectives and constraints, then outline a data-driven system that combines graph-based anomaly detection with supervised models, and finally discuss trade-offs in precision/recall, latency, and scalability. Emphasize how you would iterate from rule-based heuristics to ML models while incorporating feedback loops.

Pro tip: Frame the problem as a graph anomaly detection task and highlight the importance of temporal dynamics—fraud patterns evolve, so your system must adapt via continuous monitoring and retraining. Also, mention how you'd handle label scarcity by using weak supervision or semi-supervised learning.

1. Clarify Requirements and Define Fraud

Ask questions to understand scale, data available, latency requirements, and what constitutes self-dealing/wash transactions. Define precise metrics (e.g., precision@k, recall) and business impact.

2. Data Collection and Feature Engineering

Identify relevant data sources: transaction logs, user profiles, IP addresses, device fingerprints, reviews. Engineer features at user, seller, and graph levels (e.g., shared attributes, transaction velocity, review sentiment).

3. Modeling Approach: Graph and Sequence Models

Propose a hybrid approach: build a heterogeneous graph of users, sellers, transactions, and reviews; apply graph neural networks (GNNs) for node classification or link prediction. Complement with sequence models (LSTM/Transformer) on transaction sequences to capture temporal patterns.

4. Evaluation and Deployment

Design offline evaluation using historical labeled data and online A/B testing. Discuss deployment architecture: batch scoring for offline detection and streaming for real-time alerts. Address scalability with distributed training and inference.

5. Iteration and Feedback Loops

Incorporate human-in-the-loop feedback to label false positives/negatives, retrain models periodically, and monitor for concept drift. Suggest active learning to prioritize uncertain cases for review.

Key Points to Mention

  • Graph-based anomaly detection (e.g., GNNs, community detection) to uncover hidden relationships.
  • Temporal patterns: burstiness of transactions, review timing, and sequence modeling.
  • Handling label scarcity via semi-supervised learning, weak supervision, or anomaly detection techniques.
  • Trade-offs between precision and recall, and how to set thresholds based on business cost.
  • Scalability: distributed processing (Spark, Flink) and model serving for low-latency detection.
  • Ethical and privacy considerations: avoiding false accusations and complying with regulations.

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