Interpretable and realistic generative models in data science?
summary
A marketplace agent‑based simulator models eight interpretable parameters per product (costs, herding, five‑star bias) and applies likelihood-free Bayesian inference via sequential neural posterior estimation with normalizing flows and 1D CNN embeddings. Amortized inference trains once on simulated catalogs (≈1,400 products) and yields product‑specific posteriors, posterior predictive simulations, and about a 7% RMSE improvement on rating forecasts, demonstrating the model generates realistic time series and usable posterior uncertainty.
AI-written summary · about
chapters (6)
- 0:00Introduction
- 4:23Discriminative vs generative
- 9:58Simulation-based generative models
- 16:22Bayesian inference challenges
- 22:43Sequential neural posterior estimation
- 27:00Results and applications
related talks
33mUntangling the complexity of demand forecasting models: building a Market Simulator
01h 29mHierarchical Probabilistic Modelling in Real Life
46mWhy data-driven methods will shape the future (...)
42mSolving Real-World Business Problems with Bayesian Modeling
17mGenerative AI: Keeping it real
59mAI in Finance: Scope and Examples
38mOptimizing Retail Forecasting Through AI Models
18mKeynote: AI for Financial Services