Following up on my note from a few days ago, below is Matlab / Octave code that implements interpolated classifications. I’ve tested it on only the UCI Iris dataset, but it works, and because it’s not deterministic, the accuracy varies, and seems to range from 82% to 92%, given 350,000 Monte Carlo simulations. It’s slower than my core algorithms, but it’s still fast, and runs in 2 to 10 seconds, but starts to slow down as you increase the number of simulations.
Again, this is useful, because this algorithm does not require the data to be locally consistent (i.e., classifications can vary over small distances). This allows for classifications of datasets that might have e.g., arbitrary classifiers made up by a human, rather than a truly intrinsic classification structure.