source: distools/nnerror.m @ 69

Last change on this file since 69 was 10, checked in by bduin, 14 years ago
File size: 2.3 KB
RevLine 
[10]1%NNERROR Exact expected NN error from a dissimilarity matrix (2)
2%
3%   E = NNERROR(D,M)
4%
5% INPUT
6%   D   NxN dissimilarity dataset
7%   M   Number of objects to be selected per class (optional)
8%
9% OUTPUT
10%   E   Expected NN errror
11%
12% DEFAULT
13%   M = minimum class size minus 1
14%
15% DESCRIPTION
16% An exact computation is made of the expected NN error for a random
17% selection of M for training. D should be a dataset containing a labeled
18% square dissimilarity matrix.
19
20% Copyright: R.P.W. Duin, r.p.w.duin@prtools.org
21% Faculty EWI, Delft University of Technology
22% P.O. Box 5031, 2600 GA Delft, The Netherlands
23
24
25function e = nnerror(D,n)
26isdataset(D);
27discheck(D);
28nlab     = getnlab(D);
29[m,mm,c] = getsize(D);
30
31% Number of objects per class
32nc = classsizes(D);
33
34% Compute for all training set sizes
35if nargin < 2
36        n = min(nc)-1;
37end
38
39if length(n) > 1
40        e = zeros(1,length(n));
41        for j=1:length(n)
42                e(j) = nnerror(D,n(j));
43        end
44else
45        if n >= min(nc)
46                error('Requested size of the training set is too large.')
47        end
48        % Call for the given sample size       
49        [D,I] = sort(D);
50        I     = reshape(nlab(I),m,m);   % Order objects according to their distances
51        ee    = zeros(1,m);
52       
53        % Loop over all classes
54        for j = 1:c
55                % Find probabilities Q that objects of other classes are not selected
56                Q = ones(m,m);
57                for i = 1:c
58                        if i~=j
59                                [p,q] = nnprob(nc(i),n);
60                                q = [1,q];
61                                C = cumsum(I==i)+1;
62                                Q = Q.*reshape(q(C),m,m);
63                        end
64                end
65               
66                % Find probabilitues P for objects of this class to be the first
67                [p,q] = nnprob(nc(j),n);
68                p = [0,p];
69                C = cumsum(I==j)+1;
70                P = reshape(p(C),m,m);
71
72                % Now estimate the prob EC it is really the NN
73                J     = find(I==j);
74                EC    = zeros(m,m);
75                EC(J) = P(J).*Q(J);
76               
77                % Determine its error contribution
78                L     = find(nlab==j);
79                ee(L) = 1-sum(EC(2:m,L))./(1-EC(1,L));  % Correct for the training size
80        end
81
82        % Average for the final result
83        e = abs(mean(ee));
84end
85
86
87
88%NNPROB Probability of selection as the nearest neighbor
89%
90% [P,Q] = NNPROB(M,K)
91%
92% If K objects are selected out of M, then P(i) is the probability
93% that the i-th object is the nearest neigbor and Q(i) is the probability
94% that this object is not selected.
95
96function [p,q] = nnprob(m,k)
97p = zeros(1,m);
98q = zeros(1,m);
99q(1) = (m-k)/m;
100p(1) = k/m;
101for i=2:(m-k+1)
102        q(i) = q(i-1)*(m-k-i+1)/(m-i+1);
103        p(i) = q(i-1)*k/(m-i+1);
104end
105
Note: See TracBrowser for help on using the repository browser.