Wait Time Prediction at Yelp: Practical ML for Real-Time Estimates

Introduction

Knowing how long you'll wait at a restaurant is valuable information. Yelp's wait time prediction system provides real-time estimates using ML models trained on historical and live data.

Problem Definition

User Need

• "Should I go to this restaurant now?"
• "How busy will it be at 7 PM?"
• "Is this wait time normal?"

Prediction Task

Given:

• Restaurant characteristics
• Current time/day
• Historical patterns
• Live signals

Predict:

• Expected wait time
• Confidence interval

Data Sources

Static Features

• Restaurant type and size
• Location characteristics
• Historical average wait times

Dynamic Features

• Current time of day/week
• Weather conditions
• Local events
• Real-time check-ins

Model Architecture

Multi-Stage Approach

Input Features -> Base Prediction -> Adjustment Layer -> Final Estimate
       |               |                    |                |
   (static)     (historical avg)     (real-time)       (calibrated)

Feature Engineering

features = {
    # Time features
    'hour_of_day': encode_cyclic(hour, 24),
    'day_of_week': one_hot(day, 7),

    # Historical features
    'avg_wait_this_hour': historical_mean(restaurant, hour),
    'std_wait_this_hour': historical_std(restaurant, hour),

    # Real-time features
    'recent_checkins': count_last_hour(restaurant),
    'current_capacity': estimate_occupancy(restaurant)
}

Training Details

Data Collection

• Waitlist app data
• Check-in timestamps
• User-reported waits

Model Selection

Compared:

• Gradient boosting (XGBoost)
• Neural networks
• Time series models (Prophet)

Winner: XGBoost for interpretability and performance

Evaluation

• MAE: Mean Absolute Error
• Coverage: % within confidence interval
• Calibration: Predicted vs. actual distributions

Production System

Real-Time Updates

• New data ingested continuously
• Model features updated in real-time
• Predictions refresh on page load

Edge Cases

• New restaurants (use similar restaurant priors)
• Unusual events (increase uncertainty)
• Missing data (graceful degradation)

Results

• X minute average error on wait time
• Y% user satisfaction with estimates
• Z% increase in reservations through Yelp

Lessons Learned

1. Real-time signals matter significantly
2. Uncertainty estimates are valuable to users
3. Historical patterns provide strong baselines
4. User feedback improves predictions

---

Build real-time prediction systems with our practical ML courses.