kNN, Naive Bayes, and SVMs

Build a per-feature dissimilarity that already lives on a 0-to-1 scale and average those, which is what Gower's coefficient does: absolute difference over the feature's range for numbers, zero or one for a categorical match or mismatch.

Only redundant evidence that favours a class other than the honest winner, and whose repeated counting is large enough to overcome the honest margin. Correlated features that reinforce the class that would have won anyway just inflate confidence without changing the label.

A kernel must be symmetric and produce a positive semi-definite matrix of pairwise values on every finite set of points. Only then does it correspond to a genuine inner product in some feature space, which is what the whole method assumes.

Swap the prior, do not retrain. The class prior enters the score as one additive log term, so replacing the training log-prior with the deployment class mix corrects the model, provided the per-class feature distributions themselves have not changed.

With far more features than rows, points in general position can be separated by a hyperplane for essentially any labelling, including random ones. Perfect separation therefore carries almost no information; the width of the margin relative to the data's spread is the quantity worth reading.

Each pairwise SVM produces a signed distance measured in its own weight-norm units, learned from its own two classes. Those numbers share no common scale and are not probabilities, so comparing them across pairs is arbitrary.

Condensed nearest neighbour keeps a subset that still classifies every original training point correctly under 1-NN. It discards interior points and keeps boundary ones. The risk: mislabelled points are exactly what it preserves.

A majority of 15 needs 8 neighbours of the rare class, but in a region where that class is 4% of the data a neighbourhood of 15 typically contains none or one. The vote is therefore won by the common class everywhere, and the rare class is never predicted.