ISCA-k — Imputation and Similarity-based Classification, with adaptive k. A nearest-neighbour method for mixed-type data, and an honest map of where it works and where it does not.
Missing values are a practically unavoidable feature of real data, and the way they are handled conditions every analysis that follows. ISCA-k is a nearest-neighbour imputation method designed for mixed-type data. It addresses three limitations of the classical k-NN imputer: a fixed number of neighbours for every observation, equal weight for every variable in the proximity calculation, and reliance on a distance defined for a single variable type.
To do so, the method weights each variable by the mutual information it shares with the variable being imputed, adapts the neighbourhood size to the characteristics of each observation, and uses a generalised Gower distance while penalising pairs of observations with little shared evidence. Imputation runs sequentially, reusing values already imputed as further columns are filled.
It was compared against six reference techniques, all tuned by cross-validation, across twelve public datasets, ten missingness conditions and one hundred paired repetitions, with additional ablation studies, sensitivity analysis, explanatory hypothesis testing and cost measurement. ISCA-k obtained the best mean rank of all methods evaluated, and proved statistically equivalent to missForest at the reference condition, at roughly one fourteenth of the computational cost.
The main contribution, however, is not the method itself. It is the characterisation of its applicability frontier — the conditions under which it should be used, and those in which another method is the better choice. Those are set out below, and they are not flattering.
Nemenyi post-hoc test at α = 0.05. Methods joined by a bar are not statistically distinguishable — which is the point: ISCA-k and missForest sit in the same group, and the cost difference between them is where the argument lies.
Uniform amputation, 5% to 50%. All methods converge as information disappears; what matters is where each one stops being usable.
Seconds per imputation, logarithmic scale. The gap to missForest is the practical reason to prefer ISCA-k where accuracy is equivalent.
One-factor-at-a-time ablation. Removing the mutual-information weighting or the low-overlap penalty costs accuracy on most datasets; removing adaptive k mostly costs stability on the hard ones.
At 20% and 40% missingness. The adaptive selection rarely finds the single best k — it avoids the bad ones, which on unseen data is the more useful property.
Use it for low- to medium-dimensional datasets, particularly mixed-type ones, at missingness rates up to around 40%, and wherever the cost of missForest is prohibitive — repeated resampling, cross-validation loops, multiple imputation.
Do not use it for high-dimensional continuous data, where it loses ground clearly, or under mechanisms where missingness depends on the observed values. In that second case the assumption behind neighbour-based imputation is weakened, and the method has no way of detecting that this has happened.
Reporting this frontier is deliberate. A benchmark that only shows the conditions in which a new method wins tells the reader nothing about whether to use it on their own data. The comparison here uses six competing methods, all tuned by cross-validation rather than left at defaults, precisely because the alternative produces a flattering result that does not survive contact with practice.