Thompson sampling for personalising a car brands advertisements
summary
Claims a working system that uses Thompson sampling with hierarchical Bayesian priors to choose among 24 creative variants (2 models × 4 concepts × 3 segments), connects to a DCO platform, updates probabilities daily (no online per-impression training), and optimizes for click‑through rate. Implementation uses NumPy and custom sampling after trying PyMC3. Results show measurable uplift, clear differences between segments, and that higher‑level priors speed early learning while converging to variant‑level evidence.
AI-written summary · about
chapters (9)
- 0:00Personalised ads with Thompson sampling
- 7:33Car manufacturer case study
- 16:47Thompson sampling explained
- 23:40Bayesian approach with priors
- 30:48Experiment results and insights
- 32:12Thompson sampling vs A/B testing
- 34:31Segmentation and data sources
- 36:07Model training and context
- 37:12Tracking individuals and KPIs
related talks
35mThompson Sampling for Machine Learning
21mRise of the Machines: Removing the Human-in-the-Loop
17mByron Galbraith, Chief Data Scientist, Talla, NYC 2017
48mAdaptivity and Confounding in Multi-armed Bandit Experiments
39mSaaaS - Sampling as an Algorithm Service
42mSolving Real-World Business Problems with Bayesian Modeling
30mIs your marketing effective? Let Bayes decide!
42mA Bayesian Approach to the Cold Start Problem