Answers are grounded in this paper and cite their sources.
Exact decision-tree search is sensitive to the number of candidate predicates, motivating efficient front-end mechanisms for controlling the optimization space. We study conflict-guided feature selection (CGFS), in which discretized conflict relevance (DCR) ranks features by the fraction of sampled opposite-label pairs that they distinguish, and introduce joint conflict coverage (JCC) to reduce redundant selections. On five datasets spanning 569–45,222 instances and 14–500 original features, depth-3 PyDL8.5 experiments with a 60-s solver limit show that DCR at $$K=10$$ reduces end-to-end time by $$2.74$$ – 348.60× on completed full-feature runs, while the all-timeout Madelon configuration yields a >397.78× lower bound. Direct MurTree validation independently reproduces substantial reductions on the harder Hill–Valley and Madelon representations. JCC further recovers complementary signal groups under controlled redundancy. Together, these results establish conflict coverage as an auditable front-end criterion for reducing candidate-predicate burden before exact decision-tree optimization, with predictive effects characterized explicitly across datasets.
No takes yet. Share an insight, caveat, or question.
Li et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: