← Expedia Interview Insights

Expedia·Machine Learning Engineer·Technical Phone Screen·Senior

Senior
Apr 2026

Summary

ML system design round at Expedia where you pick a product, find the ML-worthy problems in it, and then basically walk through the entire lifecycle from problem framing to deployment. Pretty open-ended, which I was not fully prepared for.

Questions Asked (6)

Q1

Given a specific product at Expedia, how would you identify which problems within it are actually good candidates for an ML solution?

Product Sense & IdeationTechnical Trade-offs
Author's notes

I jumped straight into listing ML techniques before really justifying why ML was even needed, which felt backwards in hindsight.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by framing the answer around a specific Expedia product (e.g., hotel search ranking or trip recommendations) and then apply a structured evaluation of potential ML use cases. Focus on identifying problems where ML adds clear value over rules or heuristics, considering data availability, business impact, and feasibility.

Pro tip: Emphasize that not every problem needs ML—sometimes a simple heuristic or A/B test is better. Show you can prioritize by ROI and explain how you'd measure success with offline and online metrics.

1. Define the product and its goals

Choose a specific Expedia product (e.g., flight search, hotel recommendations) and clarify its key objectives and user pain points. This grounds the discussion in a concrete context.

2. Identify candidate problems

List potential problems within the product that could benefit from ML, such as ranking, personalization, pricing, or fraud detection. Consider where current solutions fall short.

3. Evaluate ML suitability

For each problem, assess if ML is appropriate by checking: Is there labeled data? Is the problem too complex for rules? Does it require prediction or pattern recognition? Will ML scale better than alternatives?

4. Prioritize by impact and feasibility

Rank the ML-suitable problems based on business impact (e.g., revenue, conversion, customer satisfaction) and technical feasibility (data availability, model complexity, latency).

5. Define success metrics and validation

Outline how you would measure success, including offline metrics (e.g., AUC, NDCG) and online metrics (e.g., CTR, booking rate), and plan for A/B testing.

Key Points to Mention

  • Data availability and quality: Ensure sufficient labeled data and features exist for training.
  • Business impact: Tie ML solutions to clear KPIs like conversion rate, revenue, or customer retention.
  • Technical feasibility: Consider model complexity, inference latency, and scalability.
  • Alternative solutions: Compare ML against heuristics, rules, or simpler statistical methods.
  • Ethical and fairness considerations: Avoid bias and ensure compliance with regulations.
  • Iterative approach: Start with a pilot, measure, and iterate based on feedback.

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

Q2

How would you define the ML task for this product, including inputs, outputs, and how you'd source or construct labels?

Data ModelingTechnical Trade-offs
Author's notes

Label design tripped me up.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by clarifying the product goal and the user-facing decision the model will support, then translate that into a concrete ML formulation (e.g., ranking, classification, regression). Explicitly define inputs, outputs, and label construction, and discuss trade-offs in label sourcing and potential biases.

Pro tip: Tie the label definition directly to a business metric (e.g., booking conversion) and mention how you'd validate label quality and handle feedback loops, showing you think beyond model accuracy.

1. Clarify product goal and ML objective

Ask clarifying questions to understand the product's purpose and the decision the model will inform. Map this to a standard ML task (e.g., ranking, classification, regression).

2. Define inputs and outputs

Specify the input features (e.g., user, item, context) and the output format (e.g., probability, score, class). Consider real-time vs. batch requirements.

3. Determine label sourcing strategy

Decide whether to use existing logged data, human annotation, or weak supervision. Discuss pros and cons of each, including cost, quality, and scalability.

4. Address label challenges and biases

Identify potential issues like selection bias, delayed labels, or noisy feedback. Propose mitigation strategies such as reweighting or auxiliary tasks.

5. Validate and iterate

Outline offline and online evaluation metrics aligned with business goals. Plan for monitoring and retraining to handle distribution shifts.

Key Points to Mention

  • Task type: ranking, classification, or regression based on product needs
  • Input features: user demographics, item attributes, contextual signals
  • Output: probability, score, or class with calibration considerations
  • Label sourcing: implicit feedback (clicks, bookings), explicit ratings, or human labels
  • Label quality: noise, bias, and delayed feedback
  • Evaluation: offline metrics (AUC, NDCG) and online A/B tests tied to business KPIs

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

Q3

What model class and feature signals would you choose, and why?

Technical Trade-offsSystem Design
Author's notes

Went fine.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by clarifying the business objective and available data, then propose a model class that balances performance, interpretability, and scalability. Justify your choice by linking feature signals to user behavior and travel domain specifics, and discuss trade-offs and evaluation metrics.

Pro tip: Demonstrate awareness of Expedia's two-sided marketplace by considering both traveler and supplier signals, and mention how you'd handle cold-start and seasonality—common in travel.

1. Clarify the Problem and Constraints

Ask about the specific task (e.g., ranking, recommendation, pricing), data availability, latency requirements, and business KPIs. This ensures your answer is tailored and shows you think before coding.

2. Propose a Model Class

Choose a model family (e.g., gradient boosted trees, deep neural networks, matrix factorization) and justify it based on the problem, data size, and need for interpretability or real-time inference.

3. Identify Feature Signals

List key features from user, item, context, and interaction data, such as search history, price sensitivity, destination popularity, seasonality, and supplier ratings. Explain how each signal informs the prediction.

4. Discuss Trade-offs and Evaluation

Compare your choice against alternatives in terms of accuracy, latency, scalability, and maintainability. Specify offline and online evaluation metrics (e.g., NDCG, CTR, revenue lift).

5. Address Deployment and Monitoring

Briefly cover how the model would be trained, served (batch vs. real-time), and monitored for drift, ensuring a production-ready perspective.

Key Points to Mention

  • Problem type: classification, regression, ranking, or recommendation
  • Model interpretability vs. performance trade-off (e.g., GBDT vs. deep learning)
  • Feature categories: user demographics, historical behavior, item attributes, contextual signals (time, location, device)
  • Travel-specific signals: seasonality, destination popularity, trip purpose, price elasticity
  • Cold-start strategies for new users/items using content-based or popularity features
  • Evaluation metrics: offline (AUC, NDCG) and online (A/B test, CTR, conversion rate)

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

Q4

Walk me through how you'd design the training and evaluation pipeline for this system.

System DesignTechnical Trade-offs
Author's notes

I spent too long on training and barely got to eval before they moved on.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by clarifying the system's goal and data characteristics, then outline the end-to-end pipeline from data ingestion to model deployment, emphasizing reproducibility and scalability. Structure your answer around key stages: data preparation, feature engineering, model training, evaluation, and deployment/monitoring, while discussing trade-offs at each stage.

Pro tip: Tie your design to Expedia's business context—e.g., handling large-scale travel data, personalization, and real-time inference—and mention how you'd incorporate feedback loops from online metrics to continuously improve the model.

1. Clarify Requirements and Data

Ask about the system's objective, data sources, volume, velocity, and quality. Identify if it's batch or real-time, and what business metrics matter (e.g., conversion rate, CTR).

2. Design Data Pipeline and Feature Engineering

Outline data ingestion, cleaning, transformation, and feature store integration. Discuss handling missing values, categorical encoding, and feature versioning for reproducibility.

3. Model Training and Experimentation

Describe training infrastructure (distributed training, hyperparameter tuning), model selection, and experiment tracking. Emphasize reproducibility with versioned data and code.

4. Evaluation Strategy

Define offline metrics (e.g., AUC, RMSE) and online metrics (A/B tests). Discuss validation techniques (time-based splits, cross-validation) and bias-variance trade-offs.

5. Deployment and Monitoring

Explain model serving (batch vs. real-time), CI/CD for ML, and monitoring for drift, performance, and business impact. Include retraining triggers and feedback loops.

Key Points to Mention

  • Data versioning and lineage (e.g., DVC, MLflow) for reproducibility
  • Feature store for consistency between training and serving
  • Handling class imbalance and temporal data (time-based splits)
  • Scalable training (distributed frameworks like Spark, Horovod)
  • Online evaluation via A/B testing and guardrail metrics
  • Model monitoring for data drift and concept drift, with automated retraining

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

Q5

How would you decide on online versus offline metrics, and how would you set up an A/B test to validate the model in production?

A/B Testing & ExperimentationProduct Analytics & Metrics
Author's notes

This is where things got interesting.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by framing the decision around the business objective and the trade-offs between offline metrics (fast, cheap, but imperfect proxies) and online metrics (directly measure user impact but costly and risky). Then outline a structured A/B test plan that validates the model's online performance, including hypothesis, metrics, randomization, and guardrails.

Pro tip: Emphasize that offline metrics should be used for model selection and iteration, but online metrics are the ultimate arbiter of success—and always include guardrail metrics to catch unintended negative impacts.

1. Define business objective and success criteria

Clarify what the model is trying to achieve (e.g., increase bookings, reduce cancellation) and translate that into measurable online metrics like conversion rate or revenue per user.

2. Select offline metrics as proxies

Choose offline metrics (e.g., AUC, precision@k, RMSE) that correlate with the online objective, and validate their correlation using historical A/B tests or holdout data.

3. Design the A/B test

Define hypothesis, randomization unit (e.g., user), sample size, duration, and primary/secondary/guardrail metrics. Ensure proper power analysis and avoid common pitfalls like peeking.

4. Run and monitor the experiment

Launch the test, monitor for data quality, sample ratio mismatch, and early guardrail violations. Do not stop early unless predetermined stopping rules are met.

5. Analyze results and decide

Compare online metrics between control and treatment, check statistical significance and practical significance, and decide whether to launch, iterate, or abandon based on both online and offline evidence.

Key Points to Mention

  • Offline metrics are proxies; online metrics measure true business impact.
  • Correlation between offline and online metrics should be validated before relying on offline metrics.
  • A/B test design: randomization unit, sample size, power, duration, and guardrail metrics.
  • Common pitfalls: peeking, multiple testing, novelty effects, and network effects.
  • Use of guardrail metrics to ensure no harm to user experience or business KPIs.
  • Iterative approach: use offline metrics for rapid iteration, online tests for final validation.

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

Q6

What deployment trade-offs would you consider when taking this model to production?

System DesignTechnical Trade-offs
Author's notes

Covered latency vs accuracy, batch vs real-time serving, rollback strategy.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by clarifying the model's purpose, scale, and business impact, then systematically discuss trade-offs across latency, cost, accuracy, scalability, and maintainability. Emphasize how you would measure and balance these trade-offs in the context of Expedia's travel platform, where real-time personalization and high traffic are critical.

Pro tip: Frame trade-offs in terms of business metrics (e.g., conversion rate, customer satisfaction) rather than just technical metrics, and mention how you would validate assumptions with A/B tests or shadow deployments before full rollout.

1. Clarify Requirements and Constraints

Ask about expected traffic, latency SLAs, budget, and accuracy requirements to ground the discussion in concrete numbers.

2. Identify Key Trade-off Dimensions

List dimensions such as latency vs. accuracy, cost vs. performance, scalability vs. complexity, and flexibility vs. maintainability.

3. Analyze Each Trade-off with Examples

For each dimension, explain the options (e.g., real-time vs. batch inference) and their implications for Expedia's use case.

4. Propose a Balanced Solution

Recommend a deployment strategy that optimizes for business goals, such as a hybrid approach with caching and fallbacks.

5. Outline Validation and Monitoring

Describe how you would test the deployment (A/B tests, canary releases) and monitor for drift, performance, and cost.

Key Points to Mention

  • Latency vs. accuracy trade-off: real-time inference may require model simplification or caching, while batch processing can use more complex models.
  • Cost considerations: cloud compute costs, data transfer, and storage for model artifacts and logs.
  • Scalability: handling peak travel seasons with auto-scaling and load balancing.
  • Maintainability: model versioning, rollback strategies, and CI/CD pipelines for ML.
  • Business impact: aligning deployment choices with conversion rates, customer experience, and revenue.
  • Monitoring and drift detection: tracking model performance in production and retraining triggers.

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