VDFT probs do not sum up to one

[vinodkraman]

I am trying to run VDFT on the balance-scale dataset. However, when I print the model at the end, the probabilities do not sum to 1. Also, I am noticing that under mc leaf prediction the accuracy is significantly lower compared to non mc leaf prediction. However, I'm not sure why this is the case. I've attached the code below. More specifically, its seems like the tree never splits...
code.zip

[smastelini]

Hi @vinodkraman, I did not observe the same problem.

The majority class is indeed expected to perform worse than (adaptive) Naive Bayes models. This is especially true if no splits are performed. The tree does not split because either:

• You have too few points
• There is not enough statistical evidence to enable splits

I suggest you to give a look in this book chapter for a starting point on decision trees. Also, you can refer to the guide on Hoeffding Trees that is available in the River website.

[vinodkraman]

hmm, there are only 625 examples in the dataset, how did your model get 825?

[smastelini]

I just ran your code using your data. I guess the model is using bootstrap sampling, as you are using HoeffdingAdaptiveTreeClassifier. It has a seed parameter to control the randomness of bootstrap sampling, when it is active (bootstrap_sampling=True).

[vinodkraman]

Oh shoot, totally missed this. Thanks for all the help!

[smastelini]

No worries!