← Expedia Interview Insights

Expedia·Data Scientist·Technical Phone Screen·Senior

SeniorPrefer not to say
May 2026

Summary

Deep technical interview for a Data Scientist role at Expedia focused entirely on ML evaluation and deployment for a hotel ranking model. One long multi-part question that covers basically everything from offline metrics to production monitoring. Felt more like a system design session than a typical DS interview.

Questions Asked (5)

Q1

How would you design an offline evaluation plan for a hotel ranking model that avoids data leakage, accounts for position bias, and validates that conversion probability predictions are well-calibrated?

A/B Testing & ExperimentationProduct Analytics & MetricsTechnical Trade-offs
Author's notes

This is where I spent the most time.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Structure your answer around three pillars: leakage prevention, position bias correction, and calibration validation. Start by describing a time-based split with a gap period to avoid leakage, then explain how to use inverse propensity scoring (IPS) or a position-aware model to debias evaluation, and finally detail calibration metrics like reliability diagrams and Brier score. Emphasize that these components must be integrated into a single evaluation pipeline that mirrors online A/B testing conditions.

Pro tip: Mention that you would validate the offline evaluation by comparing its rankings and calibration to a small online A/B test, ensuring the offline metrics correlate with online business metrics like conversion rate. This shows you understand the limitations of offline evaluation and the importance of iterative validation.

1. Define the evaluation goal and data split

Clarify that the goal is to estimate the model's performance in a production-like setting. Use a time-based split with a gap period (e.g., 1 day) between training and test to prevent leakage from future data, and ensure the test set reflects the deployment period.

2. Address position bias in logged data

Explain that logged data is biased because users click on higher-ranked items more often. Use inverse propensity scoring (IPS) with propensity estimated from a randomization experiment or a position-based model, or use a counterfactual evaluation method like doubly robust estimation.

3. Evaluate ranking quality with debiased metrics

Compute ranking metrics such as NDCG or MAP on the debiased data, and compare against a baseline. Ensure that the metrics account for the position bias correction, e.g., by weighting each interaction by the inverse propensity score.

4. Validate calibration of conversion probability predictions

Assess calibration using reliability diagrams, expected calibration error (ECE), and Brier score. Check calibration across different segments (e.g., hotel star rating, user country) to ensure the model is well-calibrated globally and locally.

5. Integrate and validate offline pipeline

Combine the debiased ranking and calibration evaluation into a single pipeline. Validate the offline evaluation by comparing its results to a small online A/B test, ensuring that offline improvements translate to online gains.

Key Points to Mention

  • Time-based split with a gap to prevent temporal leakage
  • Inverse propensity scoring (IPS) or position-based model for debiasing
  • Debiased ranking metrics like NDCG or MAP
  • Calibration metrics: reliability diagrams, ECE, Brier score
  • Segment-level calibration checks
  • Correlation between offline metrics and online A/B test results

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

Q2

Which ranking metrics would you report for this model, and can you explain the formulas and why they reflect actual business value rather than just model performance?

Product Analytics & MetricsTechnical Trade-offs
Author's notes

NDCG with revenue weights was my main answer, and I sketched out the formula.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by clarifying the model's objective and the business context (e.g., ranking hotels or flights for Expedia). Then, select 2-3 ranking metrics that align with business goals, explain their formulas, and discuss how they capture business value such as revenue, customer satisfaction, or conversion. Emphasize the trade-offs and why these metrics are more meaningful than generic model performance metrics.

Pro tip: Tie each metric to a specific business outcome (e.g., NDCG to booking conversion, MAP to customer engagement) and mention how you would validate the metric's impact through A/B testing or offline simulations. This shows you think beyond model evaluation to actual business impact.

1. Clarify the business objective

Ask or state the model's purpose: is it to rank hotels, flights, or activities? What is the key business goal (e.g., maximize bookings, revenue, or customer satisfaction)?

2. Select appropriate ranking metrics

Choose metrics that reflect the business goal, such as NDCG, MAP, MRR, or Precision@K. Explain why they are suitable for the ranking task.

3. Explain formulas and interpretation

Provide the mathematical formula for each metric and explain what it measures in plain terms, highlighting how it captures ranking quality.

4. Connect metrics to business value

Discuss how each metric correlates with business outcomes like conversion rate, revenue per user, or customer retention. Use examples or hypothetical scenarios.

5. Address trade-offs and validation

Mention potential trade-offs between metrics and how you would validate their business impact through A/B tests or offline simulations.

Key Points to Mention

  • NDCG (Normalized Discounted Cumulative Gain): formula, position discounting, and relevance weighting; ties to user engagement and satisfaction.
  • MAP (Mean Average Precision): formula, precision at each relevant item, averaging; reflects overall ranking quality and user click-through.
  • MRR (Mean Reciprocal Rank): formula, reciprocal of first relevant item's rank; useful when the first relevant result is critical (e.g., top hotel).
  • Precision@K and Recall@K: formulas, trade-offs, and how they relate to user attention and business constraints.
  • Business metrics: conversion rate, revenue per search, customer lifetime value, and how ranking metrics correlate with them.
  • Offline vs online evaluation: using A/B tests to measure business impact and avoid overfitting to offline metrics.

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

Q3

How would you identify the key drivers of model predictions without leaking target information into your feature analysis?

Root Cause AnalysisTechnical Trade-offs
Author's notes

SHAP came to mind immediately but I almost forgot to mention the background dataset issue, which is actually the subtle part.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by emphasizing the importance of a clean, leakage-free validation setup, such as nested cross-validation or a holdout set, to ensure feature importance is assessed on unseen data. Then describe using model-agnostic techniques like permutation importance or SHAP, computed on the validation set, to identify key drivers. Finally, discuss how to interpret and validate these drivers with domain knowledge and additional checks like partial dependence plots.

Pro tip: Always compute feature importance on a held-out validation set, not the training set, and be wary of features that are proxies for the target—check for temporal leakage by ensuring features are available at prediction time.

1. Establish a leakage-free evaluation protocol

Use a holdout set or nested cross-validation to separate data used for model training from data used for importance analysis. Ensure all feature engineering is done within cross-validation folds to avoid leakage.

2. Choose model-agnostic interpretation methods

Select techniques like permutation importance, SHAP, or LIME that can be applied to any model and computed on the validation set. These methods quantify the impact of each feature on predictions without relying on model internals.

3. Compute and rank feature importance

Calculate importance scores on the validation set, using multiple random seeds if applicable, and rank features by their average impact. Look for consistency across methods to identify robust drivers.

4. Validate drivers with domain knowledge and additional analyses

Cross-check top drivers with business logic and conduct sanity checks, such as partial dependence plots or individual conditional expectation curves, to ensure relationships make sense and are not artifacts of leakage.

5. Communicate findings and limitations

Present the key drivers with appropriate caveats about correlation vs. causation and potential leakage risks. Suggest further experiments or monitoring to confirm drivers in production.

Key Points to Mention

  • Avoid target leakage by ensuring features are available at prediction time and not derived from the target.
  • Use out-of-sample data (validation set) for feature importance to prevent overfitting and leakage.
  • Permutation importance: shuffle each feature and measure performance drop on validation set.
  • SHAP values: provide consistent, locally accurate feature attributions, computed on validation data.
  • Check for temporal leakage in time-series data by using time-based splits.
  • Validate findings with domain experts and additional interpretability tools like partial dependence plots.

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

Q4

Walk through how you'd roll out this ranking model safely, including shadow mode, canary deployments, guardrail alarms, drift detection, and an automatic rollback policy.

System DesignA/B Testing & ExperimentationTechnical Trade-offs
Author's notes

I've thought about deployment pipelines before but the late-arriving data piece caught me a bit flat.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Structure your answer as a phased rollout plan that starts with offline validation and shadow mode, then progresses to canary and full deployment with continuous monitoring. Emphasize how you define guardrail metrics, detect drift, and automate rollback to ensure safety and business impact.

Pro tip: Tie every phase to a clear exit criterion and business metric—this shows you think like a product owner, not just a modeler. Also, mention that you'd pre-register the experiment and guardrails with stakeholders to avoid post-hoc rationalization.

1. Offline Validation & Shadow Mode

First, validate the model offline on historical data and then deploy it in shadow mode to log predictions without affecting user experience. Compare its ranking against the current production model to catch discrepancies early.

2. Canary Deployment & A/B Test

Roll out to a small percentage of traffic (e.g., 1-5%) in a canary deployment, running an A/B test with guardrail metrics like latency, error rates, and business KPIs (e.g., bookings, revenue). Monitor for statistically significant regressions.

3. Guardrail Alarms & Drift Detection

Set up automated alarms for guardrail metrics (e.g., p99 latency, conversion rate) and drift detection on input features and prediction distributions. Use statistical tests (e.g., PSI, KS) to alert on significant deviations.

4. Gradual Ramp & Automatic Rollback

If canary succeeds, gradually increase traffic to 100% while continuously monitoring. Define an automatic rollback policy that triggers if guardrails are breached (e.g., conversion drops >2% or latency spikes >20%), reverting to the previous model instantly.

5. Post-Launch Monitoring & Iteration

After full rollout, maintain monitoring dashboards and periodic drift checks. Establish a feedback loop to retrain and update the model as data distributions evolve, ensuring long-term performance.

Key Points to Mention

  • Shadow mode: log predictions without impacting users, compare to production
  • Canary deployment: small traffic slice, A/B test with guardrail metrics
  • Guardrail alarms: define thresholds for latency, error rates, business KPIs
  • Drift detection: monitor feature and prediction distributions using PSI, KS, etc.
  • Automatic rollback: pre-defined triggers and instant reversion to previous model
  • Stakeholder alignment: pre-register experiment and success criteria

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

Q5

How would you verify that improving a surrogate objective offline actually translates to gains on the real business KPI, and what would you do if the offline-to-online relationship breaks down?

A/B Testing & ExperimentationProduct Analytics & MetricsAdaptability & Ambiguity
Author's notes

The offline-online gap question is one I find genuinely interesting.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by explaining how you would validate the surrogate offline through correlation and calibration with historical online experiments, then propose a small-scale online test to confirm causal impact. If the relationship breaks down, outline a diagnostic process to identify causes and adapt the surrogate or fall back to direct online measurement.

Pro tip: Emphasize the importance of pre-registering the validation criteria and maintaining a feedback loop between offline and online metrics to catch drift early. Also, mention that surrogate metrics should be leading indicators, not just correlated outcomes.

1. Validate surrogate-offline relationship

Use historical data from past online experiments to check if changes in the surrogate objective correlate with changes in the business KPI. Assess both correlation and calibration (e.g., regression slope).

2. Design a confirmatory online test

Run a small-scale A/B test or switchback experiment to measure the actual KPI impact of the intervention suggested by the surrogate. Ensure sufficient power to detect the expected effect.

3. Monitor and compare offline vs. online outcomes

During the online test, track both surrogate and KPI metrics to see if the offline predictions hold. Use sequential testing or Bayesian methods to allow early stopping if needed.

4. Diagnose breakdown causes

If the relationship breaks down, investigate potential causes: data drift, unmeasured confounders, surrogate misspecification, or changes in user behavior. Check for novelty effects or seasonality.

5. Adapt and iterate

Based on diagnosis, either refine the surrogate model, incorporate additional signals, or rely more on online experimentation. Establish a continuous validation process to prevent future breakdowns.

Key Points to Mention

  • Correlation and calibration between surrogate and KPI using historical experiments
  • Causal inference methods (e.g., instrumental variables, double ML) to strengthen offline estimates
  • A/B testing best practices: power analysis, randomization unit, guardrail metrics
  • Common reasons for offline-online mismatch: distribution shift, confounding, surrogate not capturing full causal effect
  • Iterative process: use online results to improve surrogate, and surrogate to guide online tests
  • Communication with stakeholders about uncertainty and the need for online validation

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