Multi-channel optimal path sequencing through Bayesian DL
summary
Marketing channel sequencing has ~5 trillion possible paths but only ~1 million observations, so an RNN memorizes sparse sequences and returns unstable jumps (example: flipping one channel moves predicted conversion from ~50% to ~99% then to ~0%). Extraction of RNN features into a disentangled variational latent space (two dims, beta≈4) regularizes signals, produces stable green/red clusters, and enables next-best-action suggestions. Approach is computationally expensive (three models in one) but reduces extreme probability swings. RNN + variational autoencoder handles sparsity better than plain RNN.
AI-written summary · about
chapters (5)
related talks
38mUsing Recurrent Neuronal Networks for Customer Predictions
37mImplementing Reinforcement Learning for Marketing Channel Optimization with Python
57mTutorial on Deep Learning III
01h 01mML for discrete optimization: Theoretical foundations
35mUncertainty in deep learning
34mModelling and predicting customer lifetime value
01h 30mToward Understanding In-context Learning
46mRNNs for data-driven multi-touch attribution at Booking.com