RuleFit treating integers as categorical

[exalate-issue-sync[bot]]

Rule looks like this: (AirTime in {104.0, 105.0, 110.0, 111.0, 112.0, 113.0, 114.0, 120.0, 121.0, 124.0, 125.0, 126.0, 127.0, 128.0, 130.0, 131.0, 132.0, 133.0, 136.0, 138.0, 139.0, 14.0, 140.0, 142.0, 146.0, 147.0, 148.0, 15.0, 150.0, 151.0

even though AirTime is an integer column.

Using allyears2k data for this.

It also seems like the rules may be overfitting to categoricals. It would be nice to have the option to turn on one hot encoding for this. This would also help simplify the rules.

[exalate-issue-sync[bot]]

Zuzana Olajcová commented: Hi [~accountid:557058:f0137791-c6cb-47bd-bcce-fc81ad4cfefa] , I am not able to reproduce this, can you please provide reproducible example? Thanks!

[exalate-issue-sync[bot]]

Megan Kurka commented: I don’t have the example I used anymore but still do see that it is overfitting on categorical.

[exalate-issue-sync[bot]]

Zuzana Olajcová commented: 👍 ok, I will add option to turn on one hot encoding in this JIRA. To lower the overfitting, it could also help to lower the max_num_rules.

[exalate-issue-sync[bot]]

Zuzana Olajcová commented: next steps from call with [~accountid:557058:f0137791-c6cb-47bd-bcce-fc81ad4cfefa] :

• add XGBoost algorithm and/or add option for one hot encoding in GBM/RF
• throw a clear error if min_length_rule > max_length_rule
• consolidate rules i.e.: "AGE > 5 & AGE > 10 & AGE > 12" should just be "AGE > 12"
• talk to Jean-Mattheu about the rule explainer app and see if any thing there can be incorporated in H2O-3

[exalate-issue-sync[bot]]

Zuzana Olajcová commented: moving to open as blocked by [https://h2oai.atlassian.net/browse/PUBDEV-8133|https://h2oai.atlassian.net/browse/PUBDEV-8133|smart-link]

[h2o-ops]

JIRA Issue Details

Jira Issue: PUBDEV-8011 Assignee: Zuzana Olajcová Reporter: Megan Kurka State: Open Fix Version: N/A Attachments: N/A Development PRs: N/A