Gulmammadova & Mukherjee - RNNs for data-driven multi-touch attribution at Booking.com

Explore how Booking.com uses RNNs & LSTMs for data-driven marketing attribution, featuring model architecture, experimentation strategy, challenges & results. Technical deep-dive.

Key takeaways
  • Booking.com uses Recurrent Neural Networks (RNNs), specifically LSTMs, for data-driven multi-touch attribution (DDA) modeling to understand marketing channel effectiveness

  • The model treats marketing customer journeys (MCJs) like sentences, with touchpoints as tokens and uses attention mechanisms to determine attribution weights

  • Key goals of DDA implementation:

    • Move attribution credit from lower funnel to upper funnel channels
    • Increase total conversions without harming ROI
    • Better understand marketing channel contributions
    • Make more efficient bidding decisions
  • Model architecture:

    • Embeds touchpoint features into vectors
    • Uses LSTM layers to process sequential touchpoint data
    • Applies attention mechanism to calculate attribution weights
    • Outputs conversion probability through sigmoid layer
  • Experimentation strategy:

    • Two-level experimentation approach
    • Geo-level experiments to measure total market impact
    • A/B tests to validate model performance
    • Control groups use business-as-usual attribution rules
  • Implementation challenges:

    • Lack of ground truth for attribution
    • Data quality issues with partner reporting
    • Need to balance statistical power with dilution effects
    • Complex bidding landscape across marketing channels
  • Benefits of the approach:

    • Operational simplicity
    • Easy interpretation through attention weights
    • Flexible implementation across channels
    • Ability to incorporate user characteristics
    • Compatible with existing bidding systems