PyData Global 2022

In recent years, generative models have been getting a lot of attention in the search for ML solutions that are both interpretable/explainable and robust. This is because of the hope that on the one hand, a model that generates realistic-looking datasets can also give insights into a real-world data generating process while on the other hand, such a model can be robust to outliers as it aware of the kinds of data it can or cannot generate. While deep generative models of images and text have been incredibly successful in the data generation task itself, their underlying neural network structure makes it challenging to interpret the way in which they generate data or generalize by further using them in a discriminative ML system that is also robust.

In this talk, I will talk about how realistic and detailed simulations of the problem domain might actually be the generative models we are looking for in interpretable and robust ML systems. Simulation models are common in the sciences (eg. simulations of planetary systems in astrophysics, agent-based simulations in the social sciences, neural circuit simulations in neuroscience, etc) and in engineering (eg. aerodynamic simulations in mechanical engineering, etc) and their parameters correspond directly to important causal factors in the data generating process. However, the “inverse problem” of inferring model parameters from observed data typically follows a computationally intensive grid-search like process and it is unclear how they can be used in discriminative ML tasks that dominate the world of real-life data science problems

I will go through our work on Bayesian inference in agent-based simulations of customers purchasing products on an online marketplace like Amazon.com to:
1.) Motivate the need for generative models in modern ML systems and how generative simulations have interpretability built in (~10mins).
2.) Show how recent advances in likelihood-free Bayesian inference with deep learning can infer posterior distributions over interpretable parameters at scale across the entire catalog of an online marketplace (~10mins).
3.) Use the inferred posteriors for posterior predictive simulations in rating timeseries prediction and abnormal (or fake) rating classification tasks (~10mins).

This talk will appeal to both the beginning and seasoned data science practitioner. For beginners, the first half of the talk will highlight the importance of generative models for the data science problems, show how to build such a model for an online marketplace and explain fundamental Bayesian concepts like posterior inference and posterior predictive sampling. For more experienced data scientists, the second half will go into details of likelihood-free Bayesian inference for posterior estimation and how deep probabilistic neural networks can be used for this task. Throughout the talk, I will show code snippets from our (in development) Python package to simulate and infer parameters for online marketplaces that we hope can be used by attendees directly in their own daily data science work.