Automatic Interaction Detection for GBM

[exalate-issue-sync[bot]]

For linear models, force additive models and compare to unconstrained model. For GBM, use Friedman's H-statistic.

More detail to be provided ad needed basis after further conversation with CapitalOne.

[exalate-issue-sync[bot]]

Venkatesh Yadav commented: Andy confirmed that this would be something they are interested in with Medium priority.

[exalate-issue-sync[bot]]

Arno Candel commented: Just dumping this here (WIP): {code} --- a/h2o-algos/src/main/java/hex/tree/SharedTree.java +++ b/h2o-algos/src/main/java/hex/tree/SharedTree.java @@ -475,6 +475,11 @@ public abstract class SharedTree<M extends SharedTreeModel< M,P,O>, P extends Sha DTree.Split s = dn._split; // Accumulate squared error improvements p er variable float improvement = (float)(s.pre_split_se()-s.se()); assert(improvement>=0);

• if (dn.pid()>=0) { // TODO: Only go to a certain depth
• DTree.Split ps = ((DTree.DecidedNode) _tree.node(dn.pid()))._split;
• float parentImprovement = (float) (ps.pre_split_se() - ps.se());
• Log.info("Improvement of interaction between columns " + ps.col() + " and " + s.col() + " : " + parentImprovement + improvement);
• } AtomicUtils.FloatArray.add(_improvPerVar,s.col(),improvement); } } {code}

[exalate-issue-sync[bot]]

Arno Candel commented: Similar to xgbfi, which does it via post-processing. We would want to do this during runtime, to show higher-order importances in real-time. p=2 might be enough.

[h2o-ops]

JIRA Issue Migration Info

Jira Issue: PUBDEV-3183 Assignee: Arno Candel Reporter: Venkatesh Yadav State: Open Fix Version: N/A Attachments: N/A Development PRs: N/A