Practical task are solved by complex black-box models like neural networks and decision tree ensembles today. For such type of functions it’s hard to understand, why and how decision function makes predictions, how to improve its quality, how to efficiently utialize it in production.
Those problems could be solved by representation a function in more convenient to analyse form. We will present PolyTree framework — a new way to analyse decision tree ensembles. PolyTree represents a complex GBDT ensemble in a more convenient polynomial form. This presentations has several nice properties and allows to efficiently solve many practical tasks.
We will show how PolyTree could be used to make interpretation; how it could be applied to model reduction; and how this framework allows us to represent ensembles, learned by different GBDT libraries, in one, the most computationally efficient, form.
Speaker: Vasily Ershov.
Presentation language: Russian.
Date and time: May 30th, 8:00-9:30 pm.
Location: Times, room 405.
Videos from seminars is available at http://bit.ly/MLJBSeminars