2
$\begingroup$

Sorry for a long story, but it is a long story. :)

I am using the h2o library for Python to build a decision tree and to extract the decision rules out of it. I am using some data for training where labels get TRUE and FALSE values. My final goal is to extract the significant path (leaf) of the tree where the number of TRUE cases significantly exceeds that of FALSE ones.

treemodel=H2OGradientBoostingEstimator(ntrees = 3, max_depth = maxDepth, distribution="bernoulli")
treemodel.train(x=somedata.names[1:],y=somelabel.names[0], training_frame=somedata) 
dtree = H2OTree(model = treemodel, tree_number = 0, tree_class = False)

def predict_leaf_node_assignment(self, test_data, type="Path"):        
    if not isinstance(test_data, h2o.H2OFrame): raise ValueError("test_data 
    must be an instance of H2OFrame")
       assert_is_type(type, None, Enum("Path", "Node_ID"))
    j = h2o.api("POST /3/Predictions/models/%s/frames/%s" % (self.model_id, 
    test_data.frame_id),
        data={"leaf_node_assignment": True, "leaf_node_assignment_type": 
    type})
    return h2o.get_frame(j["predictions_frame"]["name"])
dfResLabH2O.leafs = predict_leaf_node_assignment( dtree,test_data=dfResLabH2O , type="Path")      

In scikit-learn there is an option to explore the leaves by using tree.values. I understand there is no such option for h2o. Instead of that, there is an option in h2o to return predictions on leaves.

When I run dtree.predictions, I am getting pretty weird results:

dtree.predictions
Out[32]: [0.0, -0.020934915, 0.0832189, -0.0151052615, -0.13453846, -0.0039859135, 0.2931017, 0.0836743, -0.008562919, -0.12405087, -0.02181114, 0.06444048, -0.01736593, 0.13912177, 0.10727943]***

My questions (and somebody has already asked it, but no clear answer was provided so far)

  1. What's the meaning of negative predictions? I expect to get a proportions p of TRUE to ALL or FALSE to ALL, where 0<=p<=1. Is there anything wrong with my model? I ran it in scikit-learn and can point out the certain significant paths and extract rules.

  2. For positive values: is it TRUE to ALL or False to ALL proportion? I am guessing it so FALSE as I mentioned Class=False, but I am not sure.

  3. Is there any method or solution for h20 trees to reveal the sample size of the certain leaf and the [n1,n2] for TRUE and FALSE cases respectively in a similar way that scikit-learn provides?

  4. I found in some forums a function def predict_leaf_node_assignment that aims to predict on a dataset and to return the leaf node assignment (only for tree-based models), but it returns no output and I cannot find any example how to implement it.

  5. The bottom line: I'd like to be able to extract the sample size values of the leaf and to extract the specific path to it, implementing [n1,n2] or valid proportions.

I'll appreciate any kind of help and suggestions. Thank you.

$\endgroup$
3
  • $\begingroup$ Can you use code formatting to make it easier to read the code portions? Thanks! $\endgroup$
    – Wes
    Feb 12, 2019 at 19:09
  • $\begingroup$ Some more of your code might be helpful. In particular, does H2O know that your target is categorical?...It might be trying to do regression instead, leading to those negative predictions? (Edit: or, it might be reporting the log-odds rather than the probability.) $\endgroup$
    – Ben Reiniger
    Feb 13, 2019 at 2:45
  • $\begingroup$ Please see the first row, I applied GBE model for binomial family of targets. As to log odds, I calculated the proportions given that they are log odds, and I got them all between 0.46 and 0.53 - i.e all 15 leaves of the decision tree have almost equal number of TRUE and FALSE cases? Does not make sense and contradicts the sklearn findings for the same data. And generally why would non parametric Decision Tree model return log odd ratios, it's not a logistic regression... The main and the most important attribute of the tree is the proportion of the divided subsamples in the leaves $\endgroup$
    – Sapiens
    Feb 13, 2019 at 17:43

1 Answer 1

0
$\begingroup$

So far I'm not seeing a way to extract training information from the model. The H2OTree.predictions can/should give you proportion information, but won't give you leaf sample sizes. For that, you should be able to use predict_leaf_node_assignment, passing your training set in (to wastefully get passed through the model, *shrug*).

predict_leaf_node_assignment should return a dataframe with the leaf assignment for each of your training points. (The R version appears to support returning either the path or the node id, but the python one doesn't seem to have it.) You could take this, join to the original frame, and use group and aggregation functions to produce the desired [n1,n2].*

Regarding the output of predictions, see https://stackoverflow.com/questions/44735518/how-to-reproduce-the-h2o-gbm-class-probability-calculation . In particular, the default learning rate in H2O's GBM is 0.1, which helps explain your muted results.

Finally, for a little more fun with the the model's tree objects, see https://www.pavel.cool/machine%20learning/h2o-3/h2o-3-tree-api/ and https://novyden.blogspot.com/2018/12/finally-you-can-plot-h2o-decision-trees.html

*EDIT: For doing the grouping and aggregation:
(I'm more used to pandas than H2O frames, so I'll convert first. And given that H2O thinks your FALSE class is the main class, maybe those are strings not boolean?)

predicted_leaves_frame = treemodel.predict_leaf_node_assignment(data).as_data_frame()
df = data.as_data_frame()
df['binary_dep_var'] = df['dep_var'].apply(lambda x: 1 if x=='TRUE' else 0)
df['T1'] = predicted_leaves_frame['T1.C1']
print(df.groupby('T1')['binary_dep_var'].agg(['sum','count','mean'])

This should give for each leaf the number of TRUE samples and the total number of samples and the ratio. If you really need the number of FALSE samples, you could define your own aggregation function or just post-process this new dataframe.

$\endgroup$
3
  • $\begingroup$ Try it as treemodel.predict_leaf_node_assignment(data). $\endgroup$
    – Ben Reiniger
    Feb 19, 2019 at 19:26
  • $\begingroup$ OK, thanks. What I got now is a column of leaf_assignment RRLL RRRL RRRL RRRL RRRL RRRL RLRR What do I do in Python to get at least proportions if not n1 n2, for the terrminal leaves, the code I saw in your link relates to R. Thanks $\endgroup$
    – Sapiens
    Feb 19, 2019 at 20:05
  • $\begingroup$ See the new edit; I've added essentially what I had in mind with "join to the original frame, and use group and aggregation functions". (It could probably be done natively with the H2O frames.) $\endgroup$
    – Ben Reiniger
    Feb 19, 2019 at 22:07

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge that you have read and understand our privacy policy and code of conduct.

Not the answer you're looking for? Browse other questions tagged or ask your own question.