[HN Gopher] Pybaobab - Python implementation of visualization te...
___________________________________________________________________
Pybaobab - Python implementation of visualization technique for
decision trees
Author : sebg
Score : 112 points
Date : 2021-12-02 15:16 UTC (7 hours ago)
(HTM) web link (gitlab.tue.nl)
(TXT) w3m dump (gitlab.tue.nl)
| time_to_smile wrote:
| I suspect decisions trees are still highly under utilized for
| optimizing any human in the loop processes that require an actual
| script that a person needs to follow.
|
| It's not an uncommon problem that you're faced with needing to
| make a series of decisions in a business environment and old-
| school decision trees give a remarkably clear, readable output of
| the optimal way to make these decisions in as few choices as
| possible.
|
| Any data scientists working on teams with call centers, sales
| teams or customer support people would likely find a surprisingly
| useful application of this mostly forgotten (other than a
| building block for RFs) tool.
| jmmcd wrote:
| Agreed... but the decision trees we learn from data to do
| classification are quite different from the decision trees we
| design to formalise our processes. I think several replies here
| have missed the distinction.
| tsumnia wrote:
| Its not that decision trees are under utilized, but rather one
| of their disadvantages is that they can overfit to their
| training data very easily. This is actually one of the reasons
| Random Forests exist - just make a lot of overfitted decision
| trees and then vote for consensus.
|
| The visualization looks great, though it did suffer visualizing
| the Random Forest. I could see using it for a single Decision
| Tree to convey the data's structure. Definitely going to use it
| for any DT slides I have to make.
| prionassembly wrote:
| People can also follow additive scores. This is what e.g. the
| DSM does.
|
| A spreadsheet-like printed table where you mark items and sum
| scores in your head is probably easier to follow than a
| similarly-powered decision tree. Of course, you can't guarantee
| that a linear decision boundary exists, but in case there is
| one, the standard tools (Gauss-Markov/FWL theorems, p-values,
| etc.) are much, much more robust than CART or C4.5.
| __mharrison__ wrote:
| I had a weird experience recently. We tore out our deck during
| Covid lockdown and I had a rusty nail puncture my skin. The
| doctor who treated it pulled out a sheet that was essentially a
| decision tree that determined that after cleaning the wound I
| just needed to be sent home with a lollipop. (Recent tetanus
| shot was sufficient.)
|
| The strange part was when the doctor showed me the tree and
| asked me if I agreed... My response was "You're the doctor, you
| tell me?!"
| riedel wrote:
| Not really. If we are doing data science for many companies and
| explainability is an aspect we will likely go for decision
| trees if the lift for advanced models is minor anyways. We use
| this (not quite as pretty for visualization but extremely
| useful to get a grasp if the tree model):
| https://github.com/parrt/dtreeviz
| carabiner wrote:
| What I'd like is some clever way of visualizing a randomforest in
| its entirety, rather than just showing the individual trees as
| they do in
| https://gitlab.tue.nl/20040367/pybaobab/-/raw/main/images/ra....
| I have no idea how to go about doing this.
| __mharrison__ wrote:
| This is really pretty. Thanks for sharing, I will probably
| leverage this in my training!
| morelandjs wrote:
| This looks incredible! Excited to try this!
| vletal wrote:
| So was the Little Prince applying daily regularization to the
| baobabs on his planet?
| ktpsns wrote:
| Nice! For anybody who does not have access to the referenced IEEE
| paper (such as me), here it is on the website of its authors:
| https://pure.tue.nl/ws/portalfiles/portal/3522084/6724346112...
| harabat wrote:
| And the GitLab repo: https://gitlab.tue.nl/20040367/pybaobab
___________________________________________________________________
(page generated 2021-12-02 23:02 UTC)