Standard uses of PyMC3 in industry


#1

Hi all,
I get asked this question too much. So I’m writing this question here.

“We are X a tech company with Y number of users, and some funding how could we use PyMC3”.

I’ll assume for the sake of argument it’s an e-commerce or similar type of company. So trying to keep this general.

A/B testing
Generative models for understanding probability of conversion.
Bayesian survival analysis for understanding churn
Bayesian changepoint models for understanding when a distribution changes like in Dev ops analysis or when a marketing promos is launched.
Forecasting yield or ROI in advertising.
Hierarchical models for predicting pricing


#2

#3

We use a hierarchical model to predict SKU demand by week (Warehouse > Style > SKU).


#4

One approach that got mentioned was - Used to infer true work item labels when provided noisy estimates by crowd workers. Used the Dawid-Skene model.


#5

What if you’re a non-profit organization? How could we use Pymc3 there?
I happen to be in the youth development, mentoring field. I’m thinking I could use pymc3 to identify changepoints related to our policies in order to identify time of change. Perhaps I could also use pymc3 to understand when a volunteer might stop mentoring a child within the Program, although I’m not 100% sure what I would use for that…

What else could we use pymc3 for within the non-profit sector?


#6

I see no logical reasons why ‘industry’ couldn’t apply to non-profits either. The application you describe sounds a good one. You might want to use something like survival analysis for your mentoring ‘attrition’ problem. You can use PyMC3 to build those sorts of things. http://docs.pymc.io/notebooks/survival_analysis.html


#7

I recently started using Bayesian Models in a view variations (ANOVA, Hierarchical Pooling, GLM) to build a contextual anomaly detection.


#8

We A/B test the living daylight out of our application using PyMC3 and custom model diagnosis tools. We also implemented (simple) behavioural models wherever we could. For instance to understand subscription churn, separate measure of retention (probability to come back) and stickiness (how often), and a lot of other fun things!

All our analytics is based on bayesian statistics. It is daunting for most newcomers who have to read a lot when they arrive, but no one has complained so far :slight_smile: