Kernel-MCIR: Redundancy-Aware Feature Attribution by Incremental RKHS Projection Geometry

Correlated features make feature attribution ambiguous: a variable can be strongly associated with a model output while contributing little explanatory structure beyond what other variables already explain. We introduce Kernel-MCIR, a model-agnostic attribution framework that represents centered feature and output kernels in a reproducing kernel Hilbert space (RKHS) and measures how the projection of the output geometry changes when a candidate feature expands an existing conditioning subspace. This incremental projection separates the candidate's target-aligned contribution into novel and already-represented components, yielding a bounded redundancy-aware attribution score in [0,1]. Because the decomposition is order-dependent, permutation averaging provides an order-symmetrized score and treats exchangeable duplicate features symmetrically in expectation. We evaluate Kernel-MCIR across controlled redundancy, UCI Air Quality, Norwegian wind forecasting, Fashion-MNIST embeddings, and high-dimensional 20 Newsgroups representations. In the duplicate-feature stress test, the Kernel-MCIR unique component achieves a duplicate/original attribution ratio of 0.880 ± 0.114, compared with 0.965 for CKA, 0.966 for HSIC, and 0.972 for KernelSHAP. Kernel-MCIR also achieves Spearman rank stability of 1.00 on both UCI and Fashion-MNIST embeddings and approximately 0.96/0.95 under Gaussian/adversarial perturbations on wind data. Kernel-MCIR is not uniformly superior on deletion faithfulness; rather, its added value is structural: it reveals whether predictive evidence is genuinely novel or largely represented by correlated variables.