Interval fuzzy rough feature selection under simulated missingness and training label noise

Abstract

Feature selection on incomplete tables requires a criterion whose uncertainty representation is distinguished from the preprocessing used in its evaluation. We formulate an interval fuzzy-rough framework based on admissible completions and examine a simpler implementation, Interval Type-2 Fuzzy-Rough Feature Selection (IT2-FRFS), using interval Gaussian similarity scales after mean imputation. Reported experiments cover breast cancer, wine, an 800-row digits subsample, and an 800-by-20 synthetic table under 15% Missing Completely at Random (MCAR) and 10% label noise in downstream training folds, with ten corruption seeds. Selection uses clean labels. The tested midpoint-gain search reaches its feature cap on every dataset, selecting all features on wine and the synthetic table. Mean k-nearest-neighbour accuracy is 0.8738, compared with 0.8964 for minimum Redundancy Maximum Relevance (mRMR) and 0.9093 for ReliefF. A fixed-width ablation on two datasets produces small accuracy differences, without establishing an advantage of interval scales. Logistic regression improves relative to all features on breast cancer and digits, but mRMR remains stronger. The results characterize this capped, mean-imputed implementation; they do not evaluate selection from noisy labels, completion envelopes, adaptive interval widths, or missing-at-random and missing-not-at-random mechanisms. Reported pooled significance tests require reconsideration of seed dependence and validation nesting before they support general inference.