Solving White-Box Problems#
Max Kochurov
Principal Data Scientist @ PyMC Labs
My core principles are:
Be Focused. I work exclusively with interpretable models and create transparent solutions.
Be Erudite. I have wide machine learning expertise to know when you need Bayesian approach.
Be Oriented. I focus on core needs first, that is where a good project starts.
Share knowledge. I love to educate about the statistical model, so you are familiar with it.
Be Helpful. If you do not need Bayesian inference I can still help you find a research direction.
When Bayesian#
If your problem is not White Box it makes little sense to use Bayesian methods. The approach relies on clear understanding and reasoning about the problem.
The more you understand about the model and process it describes the better is the solution. No traditional approach takes this to the absolute and translates expert opinions to rigor definition.
Understanding relations is a half-way to build a Bayesian model. Ignoring these relations leads to biased estimates and decisions.
Clear interpretation is the core principle in Bayesian modeling. Every unknown parameter in the model is estimated using expert and data guidance. Using Bayesian approach you make maximum use of human in the loop data driven business processes.
If predictions is all you care about, I suggest looking into Conformal Predictions. It helps to take in account uncertainty for the prediction tasks. If this does not help, and you need uncertainty in model interpretation for your decision process, Bayesian approach is for you.
How I work#
Marketing Mix Models
Find the best allocation for your marketing budget
Brain Activity
Explain brain activity using indirect measurements
Bioinformatics
Use structural assumptions to their fullest potential
Chemistry
Iterate over complex chemical interactions to improve the whole manufacturing process
A/B Testing
Bring your critical thinking to improve the hypothesis testing pipeline
Before any project starts:
We discuss what are the real needs for you
We figure out if project worth the time investment or there is a simple workaround
We investigate how to maximize the outcome of the collaboration, split the needs to options
Worst Case
Not every project is about beautiful models. If we do not fit each other, be sure I do my best to advice you the next steps to the best of my knowledge.
Teaching#
Here are my materials that may interest you:
I created modern Practical Bayes course for MSU. It focuses on practical interpretable analytics and model analysis.
The course is recorded on YouTube in State of Bayes playlist.
For more learning materials follow the link below:
Talks#
I regularly participate in conferences, give talks and share knowledge about Bayesian statistics
I link all the available materials:
Video
Presentation
Code