Daniel@0: 
Daniel@0: %SOM_DEMO4 Data analysis using the SOM.
Daniel@0: 
Daniel@0: % Contributed to SOM Toolbox 2.0, February 11th, 2000 by Juha Vesanto
Daniel@0: % http://www.cis.hut.fi/projects/somtoolbox/
Daniel@0: 
Daniel@0: % Version 1.0beta juuso 071197 
Daniel@0: % Version 2.0beta juuso 090200 070600
Daniel@0: 
Daniel@0: clf reset;
Daniel@0: f0 = gcf;
Daniel@0: echo on
Daniel@0: 
Daniel@0: 
Daniel@0: 
Daniel@0: 
Daniel@0: clc
Daniel@0: %    ==========================================================
Daniel@0: %    SOM_DEMO4 - DATA ANALYSIS USING THE SOM
Daniel@0: %    ==========================================================
Daniel@0: 
Daniel@0: %    In this demo, the IRIS data set is analysed using SOM. First, the
Daniel@0: %    data is read from ascii file (please make sure they can be found
Daniel@0: %    from the current path), normalized, and a map is
Daniel@0: %    trained. Since the data also had labels, the map is labelled.
Daniel@0: 
Daniel@0: try, 
Daniel@0:   sD = som_read_data('iris.data');     
Daniel@0: catch
Daniel@0:   echo off
Daniel@0: 
Daniel@0:   warning('File ''iris.data'' not found. Using simulated data instead.')
Daniel@0:   
Daniel@0:   D = randn(50,4); 
Daniel@0:   D(:,1) = D(:,1)+5;     D(:,2) = D(:,2)+3.5; 
Daniel@0:   D(:,3) = D(:,3)/2+1.5; D(:,4) = D(:,4)/2+0.3;
Daniel@0:   D2 = randn(100,4); D2(:,2) = sort(D2(:,2));
Daniel@0:   D2(:,1) = D2(:,1)+6.5; D2(:,2) = D2(:,2)+2.8; 
Daniel@0:   D2(:,3) = D2(:,3)+5;   D2(:,4) = D2(:,4)/2+1.5;  
Daniel@0:   sD = som_data_struct([D; D2],'name','iris (simulated)',...
Daniel@0: 		       'comp_names',{'SepalL','SepalW','PetalL','PetalW'});
Daniel@0:   sD = som_label(sD,'add',[1:50]','Setosa');
Daniel@0:   sD = som_label(sD,'add',[51:100]','Versicolor');
Daniel@0:   sD = som_label(sD,'add',[101:150]','Virginica');
Daniel@0:   
Daniel@0:   echo on
Daniel@0: end
Daniel@0: 
Daniel@0: sD = som_normalize(sD,'var');
Daniel@0: sM = som_make(sD);
Daniel@0: sM = som_autolabel(sM,sD,'vote');
Daniel@0: 
Daniel@0: pause % Strike any key to visualize the map...
Daniel@0: 
Daniel@0: clc 
Daniel@0: %    VISUAL INSPECTION OF THE MAP
Daniel@0: %    ============================
Daniel@0: 
Daniel@0: %    The first step in the analysis of the map is visual inspection.
Daniel@0: %    Here is the U-matrix, component planes and labels (you may
Daniel@0: %    need to enlarge the figure in order to make out the labels).
Daniel@0: 
Daniel@0: som_show(sM,'umat','all','comp',[1:4],'empty','Labels','norm','d');
Daniel@0: som_show_add('label',sM.labels,'textsize',8,'textcolor','r','subplot',6);
Daniel@0: 
Daniel@0: %    From this first visualization, one can see that:
Daniel@0: %     - there are essentially two clusters
Daniel@0: %     - PetalL and PetalW are highly correlated
Daniel@0: %     - SepalL is somewhat correlated to PetalL and PetalW
Daniel@0: %     - one cluster corresponds to the Setosa species and exhibits
Daniel@0: %       small petals and short but wide sepals
Daniel@0: %     - the other cluster corresponds to Virginica and Versicolor
Daniel@0: %       such that Versicolor has smaller leaves (both sepal and
Daniel@0: %       petal) than Virginica
Daniel@0: %     - inside both clusters, SepalL and SepalW are highly correlated
Daniel@0: 
Daniel@0: pause % Strike any key to continue...
Daniel@0: 
Daniel@0: %    Next, the projection of the data set is investigated. A
Daniel@0: %    principle component projection is made for the data, and applied
Daniel@0: %    to the map. The colormap is done by spreading a colormap on the
Daniel@0: %    projection. Distance matrix information is extracted from the
Daniel@0: %    U-matrix, and it is modified by knowledge of zero-hits
Daniel@0: %    (interpolative) units. Finally, three visualizations are shown:
Daniel@0: %    the color code, with clustering information and the number of
Daniel@0: %    hits in each unit, the projection and the labels.
Daniel@0: 
Daniel@0: echo off
Daniel@0: 
Daniel@0: f1=figure;
Daniel@0: [Pd,V,me,l] = pcaproj(sD,2); Pm = pcaproj(sM,V,me); % PC-projection
Daniel@0: Code = som_colorcode(Pm); % color coding
Daniel@0: hits = som_hits(sM,sD);  % hits
Daniel@0: U = som_umat(sM); % U-matrix
Daniel@0: Dm = U(1:2:size(U,1),1:2:size(U,2)); % distance matrix
Daniel@0: Dm = 1-Dm(:)/max(Dm(:)); Dm(find(hits==0)) = 0; % clustering info
Daniel@0: 
Daniel@0: subplot(1,3,1)
Daniel@0: som_cplane(sM,Code,Dm);
Daniel@0: hold on
Daniel@0: som_grid(sM,'Label',cellstr(int2str(hits)),...
Daniel@0: 	 'Line','none','Marker','none','Labelcolor','k');
Daniel@0: hold off 
Daniel@0: title('Color code')
Daniel@0: 
Daniel@0: subplot(1,3,2)
Daniel@0: som_grid(sM,'Coord',Pm,'MarkerColor',Code,'Linecolor','k');
Daniel@0: hold on, plot(Pd(:,1),Pd(:,2),'k+'), hold off, axis tight, axis equal
Daniel@0: title('PC projection')
Daniel@0: 
Daniel@0: subplot(1,3,3)
Daniel@0: som_cplane(sM,'none')
Daniel@0: hold on
Daniel@0: som_grid(sM,'Label',sM.labels,'Labelsize',8,...
Daniel@0: 	 'Line','none','Marker','none','Labelcolor','r');
Daniel@0: hold off
Daniel@0: title('Labels')
Daniel@0: 
Daniel@0: echo on
Daniel@0: 
Daniel@0: %    From these figures one can see that: 
Daniel@0: %     - the projection confirms the existence of two different clusters
Daniel@0: %     - interpolative units seem to divide the Virginica
Daniel@0: %       flowers into two classes, the difference being in the size of
Daniel@0: %       sepal leaves
Daniel@0:     
Daniel@0: pause % Strike any key to continue...
Daniel@0: 
Daniel@0: %    Finally, perhaps the most informative figure of all: simple
Daniel@0: %    scatter plots and histograms of all variables. The species
Daniel@0: %    information is coded as a fifth variable: 1 for Setosa, 2 for
Daniel@0: %    Versicolor and 3 for Virginica. Original data points are in the
Daniel@0: %    upper triangle, map prototype values on the lower triangle, and
Daniel@0: %    histograms on the diagonal: black for the data set and red for
Daniel@0: %    the map prototype values. The color coding of the data samples
Daniel@0: %    has been copied from the map (from the BMU of each sample). Note
Daniel@0: %    that the variable values have been denormalized.
Daniel@0: 
Daniel@0: echo off
Daniel@0: 
Daniel@0: % denormalize and add species information
Daniel@0: names = sD.comp_names; names{end+1} = 'species';
Daniel@0: D = som_denormalize(sD.data,sD); dlen = size(D,1);
Daniel@0: s = zeros(dlen,1)+NaN; s(strcmp(sD.labels,'Setosa'))=1;
Daniel@0: s(strcmp(sD.labels,'Versicolor'))=2; s(strcmp(sD.labels,'Virginica'))=3;
Daniel@0: D = [D, s];
Daniel@0: M = som_denormalize(sM.codebook,sM); munits = size(M,1);
Daniel@0: s = zeros(munits,1)+NaN; s(strcmp(sM.labels,'Setosa'))=1;
Daniel@0: s(strcmp(sM.labels,'Versicolor'))=2; s(strcmp(sM.labels,'Virginica'))=3;
Daniel@0: M = [M, s];
Daniel@0: 
Daniel@0: f2=figure;
Daniel@0: 
Daniel@0: % color coding copied from the map
Daniel@0: bmus = som_bmus(sM,sD); Code_data = Code(bmus,:); 
Daniel@0: 
Daniel@0: k=1; 
Daniel@0: for i=1:5, for j=1:5, 
Daniel@0:     if i<j, i1=i; i2=j; else i1=j; i2=i; end
Daniel@0:     subplot(5,5,k); cla
Daniel@0:     if i<j,
Daniel@0:       som_grid('rect',[dlen 1],'coord',D(:,[i1 i2]),...
Daniel@0: 	       'Line','none','MarkerColor',Code_data,'Markersize',2);
Daniel@0:       title(sprintf('%s vs. %s',names{i1},names{i2}))
Daniel@0:     elseif i>j,
Daniel@0:       som_grid(sM,'coord',M(:,[i1 i2]),...
Daniel@0: 	       'markersize',2,'MarkerColor',Code);
Daniel@0:       title(sprintf('%s vs. %s',names{i1},names{i2}))
Daniel@0:     else
Daniel@0:       if i1<5, b = 10; else b = 3; end
Daniel@0:       [nd,x] = hist(D(:,i1),b); nd=nd/sum(nd); 
Daniel@0:       nm = hist(M(:,i1),x); nm = nm/sum(nm);
Daniel@0:       h=bar(x,nd,0.8); set(h,'EdgeColor','none','FaceColor','k'); 
Daniel@0:       hold on 
Daniel@0:       h=bar(x,nm,0.3); set(h,'EdgeColor','none','FaceColor','r'); 
Daniel@0:       hold off
Daniel@0:       title(names{i1})
Daniel@0:     end
Daniel@0:     k=k+1;
Daniel@0:   end
Daniel@0: end
Daniel@0: 
Daniel@0: echo on
Daniel@0: 
Daniel@0: %    This visualization shows quite a lot of information:
Daniel@0: %    distributions of single and pairs of variables both in the data
Daniel@0: %    and in the map. If the number of variables was even slightly
Daniel@0: %    more, it would require a really big display to be convenient to
Daniel@0: %    use.
Daniel@0: 
Daniel@0: %    From this visualization we can conform many of the earlier
Daniel@0: %    conclusions, for example: 
Daniel@0: %     - there are two clusters: 'Setosa' (blue, dark green) and 
Daniel@0: %       'Virginica'/'Versicolor' (light green, yellow, reds)
Daniel@0: %       (see almost any of the subplots)
Daniel@0: %     - PetalL and PetalW have a high linear correlation (see
Daniel@0: %       subplots 4,3 and 3,4)
Daniel@0: %     - SepalL is correlated (at least in the bigger cluster) with
Daniel@0: %       PetalL and PetalW (in subplots 1,3 1,4 3,1 and 4,1)
Daniel@0: %     - SepalL and SepalW have a clear linear correlation, but it
Daniel@0: %       is slightly different for the two main clusters (in
Daniel@0: %       subplots 2,1 and 1,2)       
Daniel@0: 
Daniel@0: pause % Strike any key to cluster the map...
Daniel@0: 
Daniel@0: close(f1), close(f2), figure(f0), clf
Daniel@0: 
Daniel@0: clc 
Daniel@0: %    CLUSTERING OF THE MAP
Daniel@0: %    =====================
Daniel@0: 
Daniel@0: %    Visual inspection already hinted that there are at least two
Daniel@0: %    clusters in the data, and that the properties of the clusters are
Daniel@0: %    different from each other (esp. relation of SepalL and
Daniel@0: %    SepalW). For further investigation, the map needs to be
Daniel@0: %    partitioned.
Daniel@0: 
Daniel@0: %    Here, the KMEANS_CLUSTERS function is used to find an initial
Daniel@0: %    partitioning. The plot shows the Davies-Boulding clustering
Daniel@0: %    index, which is minimized with best clustering.
Daniel@0: 
Daniel@0: subplot(1,3,1)
Daniel@0: [c,p,err,ind] = kmeans_clusters(sM, 7); % find at most 7 clusters
Daniel@0: plot(1:length(ind),ind,'x-')
Daniel@0: [dummy,i] = min(ind)
Daniel@0: cl = p{i};
Daniel@0: 
Daniel@0: %    The Davies-Boulding index seems to indicate that there are
Daniel@0: %    two clusters on the map. Here is the clustering info
Daniel@0: %    calculated previously and the partitioning result: 
Daniel@0: 
Daniel@0: subplot(1,3,2)
Daniel@0: som_cplane(sM,Code,Dm)
Daniel@0: subplot(1,3,3)
Daniel@0: som_cplane(sM,cl)
Daniel@0: 
Daniel@0: %    You could use also function SOM_SELECT to manually make or modify
Daniel@0: %    the partitioning.
Daniel@0: 
Daniel@0: %    After this, the analysis would proceed with summarization of the
Daniel@0: %    results, and analysis of each cluster one at a time.
Daniel@0: %    Unfortunately, you have to do that yourself. The SOM Toolbox does
Daniel@0: %    not, yet, have functions for those purposes.
Daniel@0: 
Daniel@0: pause % Strike any key to continue...
Daniel@0: 
Daniel@0: 
Daniel@0: clf
Daniel@0: clc 
Daniel@0: %    MODELING
Daniel@0: %    ========
Daniel@0: 
Daniel@0: %    One can also build models on top of the SOM. Typically, these
Daniel@0: %    models are simple local or nearest-neighbor models. 
Daniel@0: 
Daniel@0: %    Here, SOM is used for probability density estimation. Each map 
Daniel@0: %    prototype is the center of a gaussian kernel, the parameters
Daniel@0: %    of which are estimated from the data. The gaussian mixture
Daniel@0: %    model is estimated with function SOM_ESTIMATE_GMM and the
Daniel@0: %    probabilities can be calculated with SOM_PROBABILITY_GMM.
Daniel@0: 
Daniel@0: [K,P] = som_estimate_gmm(sM,sD);
Daniel@0: [pd,Pdm,pmd] = som_probability_gmm(sD,sM,K,P);
Daniel@0: 
Daniel@0: %    Here is the probability density function value for the first data
Daniel@0: %    sample (x=sD.data(:,1)) in terms of each map unit (m):
Daniel@0: 
Daniel@0: som_cplane(sM,Pdm(:,1))
Daniel@0: colorbar
Daniel@0: title('p(x|m)')
Daniel@0: 
Daniel@0: pause % Strike any key to continue...
Daniel@0: 
Daniel@0: %    Here, SOM is used for classification. Although the SOM can be
Daniel@0: %    used for classification as such, one has to remember that it does
Daniel@0: %    not utilize class information at all, and thus its results are
Daniel@0: %    inherently suboptimal. However, with small modifications, the
Daniel@0: %    network can take the class into account. The function
Daniel@0: %    SOM_SUPERVISED does this.
Daniel@0: 
Daniel@0: %    Learning vector quantization (LVQ) is an algorithm that is very
Daniel@0: %    similar to the SOM in many aspects. However, it is specifically
Daniel@0: %    designed for classification. In the SOM Toolbox, there are
Daniel@0: %    functions LVQ1 and LVQ3 that implement two versions of this
Daniel@0: %    algorithm.
Daniel@0: 
Daniel@0: %    Here, the function SOM_SUPERVISED is used to create a classifier
Daniel@0: %    for IRIS data set:
Daniel@0: 
Daniel@0: sM = som_supervised(sD,'small');
Daniel@0: 
Daniel@0: som_show(sM,'umat','all');
Daniel@0: som_show_add('label',sM.labels,'TextSize',8,'TextColor','r')
Daniel@0: 
Daniel@0: sD2 = som_label(sD,'clear','all'); 
Daniel@0: sD2 = som_autolabel(sD2,sM);       % classification
Daniel@0: ok = strcmp(sD2.labels,sD.labels); % errors
Daniel@0: 100*(1-sum(ok)/length(ok))         % error percentage (%)
Daniel@0: 
Daniel@0: echo off
Daniel@0: 
Daniel@0: 
Daniel@0: 
Daniel@0: 
Daniel@0: 
Daniel@0: 
Daniel@0: