Overview
A conference-ready study comparing Naive Bayes instance filtering and decision-tree attribute selection/weighting for noise handling in DT-NB hybrids, run separately and in both orders. The authoritative experiment uses Weka J48 3.8.6 and FaithfulNB over the original attribute columns with seeds 0-9 and stratified 10-fold cross-validation, refitting every supervised step inside each training fold and scoring on held-out rows. Plain C4.5 averaged 86.33% accuracy against 81.21% (instances-first) and 80.89% (attributes-first), and none of the eight post-hoc comparisons survived Holm correction, framed honestly as a protocol-sensitivity study rather than a win claim. Results are archived with hashes and the repo carries three manuscript variants. The manuscript is written but has not been submitted to any conference.