Stack Decision Region Plots in 3D

[gfcataldo]

Hi everyone. First of all, I'm quite new at using matplotlib and mlxtend, so I apologise in advance if what I ask is stupid or simply wrong.

The thing is that I want to stack some decision region plots in 3D depending on the filler value that I set for the third feature so it looks similar to this:

From: https://kaleidoscopicdiaries.wordpress.com/2016/08/22/stacking-density-plots-in-3d-in-matplotlib/

Since the plot_decision_regions method returns a matplotlib.axes.Axes object, I have not been able to figure out how to do it properly. Is there an easy way to do it?

[rasbt]

Hm, this sounds interesting but tricky. You can actually pass an axes object to the function, which you would need for a 3D projection, i.e., sth like

ax = fig.add_subplot(111, projection='3d')

However, the parts

 ax.contourf(X, Y, A1, 100, zdir='z', offset=0)
ax.contourf(X, Y, A2, 100, zdir='z', offset=0.5)
ax.contourf(X, Y, A3, 100, zdir='z', offset=1)

is where it would become tricky, as there's currently no way of doing that via the input arguments. I think your best bet would be modifying the function for your purposes:
https://github.com/rasbt/mlxtend/blob/master/mlxtend/plotting/decision_regions.py

or just return the xx, yy, Z from line 234 (https://github.com/rasbt/mlxtend/blob/master/mlxtend/plotting/decision_regions.py#L234) and collect them from multiple filler value runs and then pass them to your custom contourf 3D plots

[gfcataldo]

Thank you very much, I'll give it a try.

Also, another quick question. When I plot the decision regions of a decision tree, the function returns something like this:

My question is: which category do the figures that are in the edges fall into? Or, how do I know if the algorithm predicts they are a 0 or a 1? Because I was expecting the edges to fall in decimal values, so it would be easier to identify those cases. By the way, all the values I work with are integers.

[rasbt]

Hard to tell. If you are using scikit-learn decision trees, they set the threshold at x_i <= x_j. So if your the red squares are at x_1 = 2 are on the boundary, they will fall left of the boundary. The line then does not appear precisely at the tickmark for x_1 = 2, but that may be due to

a) imprecision of the way it visually appears when converted to a given file format like png
b) the resolution of the grid along the x_1 axis is not large enough and it may only evaluate points like ... 1.99, 2.01 ... In this case, you can address this by choosing a smaller value for the res parameter in teh decision plot function but it will take more computational time

[rasbt]

Based on how complicated the function has become recently, this should really be a plot_decision_regions_3d to not further complicate things.