KNN+X is a proposed improvement to the well-known KNN algorithm.
To label a new sample, the k nearest samples from the training data are found
and then another classification algorithm X is used on the k neighbors.
This is a replacement for the original majority choice in KNN, this is meant to make a more informed choice in cases where the majority is slight.

We enhance the proposed method by adding thresholds. Using thresholds, we can decide that if the majority label is 80% of the k-neighborhood, we choose it (t=0.8 in code).
So if we have a sufficient majority we don't train a new classifier - this saves runtime.

The experiment python code requires:

• numpy
• pandas
• h5py
• sklearn
• joblib

The plotting notebook also requires Matplotlib.

When running experiments you can control various variables:

• Runs amount - how many experiment rounds to run.
• K values - which values of k to experiment with. e.g. 5, 13, 20, 100. These should be positive integers
• t values - which values of t to experiment with. e.g. 0.2,0.6,0.9,1. We require 0 <= t <= 1.
• Datasets - which datasets to experiment with (Currently implemented: covertype, glass, mnist, skin, shuttle, usps, wine and yeast).
• KNN algo - which KNN algorithm to use. (Choose between brute, kd_tree and ball_tree).

To run the experiment code, optionally choose the run amount, datasets, threshold values, k values, and knn algorithm to try in the main function in knn_plus_x.py and run it.
python knn_plus_x.py

You can plot the results using plot_results.ipynb