sofm.m

clear; close all; clc

%% Performing clustering with only "Self-organizing Maps"
% Self-organizing maps learn to cluster data based on similarity, 
% topology, with a preference (but no guarantee) of assigning the same 
% number of instances to each class. Self-organizing maps are used both to 
% cluster data and to reduce the dimensionality of data. (from MATLAB help)
%
% By the word "only", we mean that all the clustering process is achieved
% by only creating a SOFM neural network. For this purpose, the number of
% the desired clusters, must be the same as the number of neurons in the
% output layer of the SOFM neural network.

%% Part 1- Creating a fake dataset
nRow_fake_dataset = 3000; % it should actually be 30000, but for elapsed 
% running time of the code, 3000 is chosen.  
nCol_fake_dataset = 600; % it should actually be 6000, but for elapsed 
% running time of the code, 600 is chosen.

temp0 = randn(ceil(nRow_fake_dataset/5), ceil(nCol_fake_dataset/5));
temp1 = repmat(temp0, 5, 5);
fake_dataset = temp1(1:nRow_fake_dataset, 1:nCol_fake_dataset);

%% Part 2- Performing linear dimensionality reduction (or noise suppression) alg.
[coeff, score, ~, ~, explained, ~] = pca(fake_dataset);
explained_variance_threshold = 99.5; % for the FFT dataset, a threshold of 
% 99 or 99.5 is thought to be more suitable.

cumulative_explained = cumsum(explained);
numFeatures = find(cumulative_explained >= explained_variance_threshold);
numFeatures = numFeatures(1);

reduced_dataset_by_pca = score(:, 1:numFeatures);
size(reduced_dataset_by_pca)

%% Part 3- Performing SOFM (a non-linear and trainable dimensionality reduction alg.)
%% Part 3-1- Preparing (transposing) the dataset matrix
inputs = reduced_dataset_by_pca.';

%% Part 3-2- Creating a "network" object for SOFM
numClusters = 3; 
row_nodes = 3; column_nodes = 1;
% Note: the result of the product "row_nodes * column_nodes", should be
% equal to "numClusters".

dimensions = [row_nodes, column_nodes];
net = selforgmap(dimensions);
net = configure(net, inputs);
figure; plotsompos(net, inputs); grid on

%% Part 3-3- Training SOFM and visualizing its performance benchmarks 
[net, tr] = train(net, inputs, ...
    'CheckpointFile', 'sofm_chk', 'CheckpointDelay', 2);

view(net)
figure; plotsompos(net, inputs); grid on
figure; plotsomhits(net, inputs);
figure; plotsomnd(net);

%% Part 3-4- Clustering dataset based on the trained SOFM
oneHotClasses = net(inputs);
classes = vec2ind(oneHotClasses);

%% Part 3-5- Visualization of the cluster proposals for some samples
numOfSamples = 5;
disp(['Proposals for the first ' num2str(numOfSamples) ' samples:'])
classes(1:numOfSamples)