RNNs for data-driven multi-touch attribution at Booking.com
summary
Builds an attention-based LSTM that predicts conversion from customer journeys, using co-trained embeddings for categorical features and a softmax attention layer whose weights sum to one and serve as per-touchpoint credit (example outputs 45/35/20 on three clicks). Adds a tiled start-touch user feature to capture loyalty, uses credit-shift analysis to check that credit moves from last-touch to upper-funnel channels, tolerates ~10% data reporting mismatch, and runs a search-destination A/B (with a planned geo-level followup); first-step experiments complete and results are under assessment.
AI-written summary · about
chapters (13)
- 0:00Introduction
- 1:07Speakers and project
- 4:18Marketing attribution challenges
- 6:02Customer journey data
- 8:11Single-touch limits
- 9:14Multi-touch attribution
- 12:45RNNs with attention
- 20:35User characteristics in model
- 23:04Offline evaluation
- 25:52Online experiment design
- 31:16Randomization unit
- 36:14Two-step framework
- 37:41Q&A
related talks
38mUsing Recurrent Neuronal Networks for Customer Predictions
37mMulti-touch Attribution: What am I training for?
33mDeep Learning at Booking.com
21mHow to Get Ahead in Advertising
33mCausal Inference Framework for incrementality
30mWinning (Some) Churners Back
24mUnlocking revenue: newcomer classification for boosted profits
30mMobile Order Click-Through Rate (CTR) Recommendation with Ray on Apache Spark at Burger King