Using Recurrent Neuronal Networks for Customer Predictions
summary
RNNs (LSTM/GRU-style cells) are trained on 2.4M sessions spanning 180 days to predict whether a session converts, using embeddings plus time-between-events as inputs and a sequence length ~150; model has ~200k weights, is trained 250 passes (~20 hours) on Tesla GPUs and deployed to score live sessions and push segments to DSPs. Results: sweet-spot recall ~35% (kitchen site ~50%), CTR +20% and conversion uplift from 0.12% to 0.26%; 2.4M sessions, 180 days trained 250 epochs, ~20 hours on Tesla GPUs 200k weights and sequence length ~150
AI-written summary · about
chapters (23)
- 0:00Introduction
- 1:40What is Machine Learning
- 6:16Decision Tree Example
- 9:29Neuronal Networks
- 12:16Experiment
- 13:32Training Data
- 14:38Embedding
- 15:44RNA
- 16:54Input and Output
- 17:35Internal State
- 19:17User Journey
- 20:14Case Studies
- 20:59Data Preparation
- 21:59Training
- 22:57Tesla GPU
- 24:09Binary Classification
- 24:40Metrics
- 26:09Thresholds
- 27:36Clickthrough rate
- 28:50Running the network differently
- 29:45Training the network on jokes
- 30:13Wrap up
- 31:17QA
related talks
27mMore Then Words : Business Applications of Recurrent Networks
46mRNNs for data-driven multi-touch attribution at Booking.com
13mBeyond Surveys: Real-Time User Satisfaction Prediction
26mImproving customer experience through journey...
31mLarge Scale CTR Prediction Lessons Learned
49mUnderstanding Customer Journeys with Deep Learning
31mThe Advantages of Deep Learning with Recurrent Neural Networks
43mPersonalization and Scalable Deep Learning with