Trees of the ensemble present unnecessary splits

[lorebon]

Dear Isak,

during my experiments I've encountered some ShapeletTrees that present meaningless splits, in the sense that 2 leaf nodes deriving from a certain branching node have the same classification label.

Is there a reason for that? It seems that the tree wants to reach the max depth regardless. I'm sure this does not impact performance, but I still wonder why such behavior.

Thanks!

[isaksamsten]

It sounds like a bug in the stopping criteria. Any way for me to reproduce?

[isaksamsten]

Could you try setting min_impurity_decrease to a value slightly larger than 0 and see if the issue persists? Could be some minor numerical deviation where the impurity should be zero but is very slightly above du to numerical instability.

[lorebon]

Ok I'm trying with min_impurity_decrease equal to 1e-3 and it seems working! Thank you!

[isaksamsten]

Thanks for the report. I will push a small change to fix the issue.

[lorebon]

I'm still facing the same issue with a value of 0.08.

Here is the minimal reproducible example I can provide:

test_min_decrease.txt

If you run it you can notice that leaf nodes 4 and 5 have the same label (1).

With min_impurity_decrease above 0.08 there is only the root node.

[isaksamsten]

The attribute value encodes the probability distribution over the labels reaching the node. So in this case 18 samples reaches the node (you can use the attribute n_node_samples to confirm this) out of which 8 are labeled as -1 and 8 as 1.

[lorebon]

I've just checked and nodes 4 and 5 present the following probabilities:
4) 0.5333 0.4666
5) 0.8857 0.1143
So the np.argmax is rightfully choosing label 1 for both nodes.

[isaksamsten]

What version are you using? Just so I test the right version

[lorebon]

I tried with both 1.1.2 and 1.1.3

[isaksamsten]

I have a hypothesis on what can go wrong. here we skip shapelets that produce constant distances to all samples. The bug would present it self if all shapelets produce such constant distances and the best split is not initialised. Perhaps that’s what’s happening. I will try to implement some sanity check for that case and see if I can reproduce the bug.

[isaksamsten]

Hi! I will take a look! Thanks for reporting