This repository contains a project using k-NN (k-Nearest Neighbors) classifier. As example, five datasets are included in the project i.e. the famous iris, e-coli proteins, yeasts proteins, wine chemistry and weat seeds.
There are 3 python modules, which are explained next. Also, there is a folder called data_raw containing one folder for each of the five datasets. This folder contains the raw data. The foldar data stores the subsets i.e. validation (Z2), test (Z3) and training (Z1) sets.
This module receives the raw information contained in the data_raw folder, and randomly splits the samples into data folder, this way generating the validation (Z2), test (Z3) and training (Z1) sets.
This module is imported in the methods.py module. The functions are used in two training methods provided by the methods.py module, which are explained in the the next module.
In this, there are two methods to train the classifier.
First, the validation set is classified on the training for a predefined number of iterations (predefined: 30), varying the parameter k from the k-NN classifier. After, the value for k that minimize the errors is considered and the misclassified samples on the validation are swaped with random sampled from the training set, this procedure repeats for a predefined number of runs (predefined: 30). Last, the test set is classified on the optimized training set to evaluate the generated classification model.
This method is the same as method 1, the only difference is on training step. In method 1, when the runs occur swapping the samples the k is fixed. In this method, the is also optimized again when the runs occur.