1 | %STRKERM String Kernel Matrix by Lodhi et al
|
---|
2 | %
|
---|
3 | % [K,KK] = STRKERM (A,B,W,LAM,KNORM)
|
---|
4 | %
|
---|
5 | % INPUT
|
---|
6 | % A Cell structure of N strings or a single string
|
---|
7 | % B Cell structure of M strings or a single string (optional; default: B = A)
|
---|
8 | % W Vector of K weights (optional; default: [1 1])
|
---|
9 | % LAM Decay factor Weight parameter in (0,1] used to scale the intermediate kernels
|
---|
10 | % (optional; default: 1)
|
---|
11 | % KNORM Parameter (0/1) indicating whether the kernel should be normalized
|
---|
12 | % to remove bias wrt to text length (optional; default: 1)
|
---|
13 | %
|
---|
14 | % OUTPUT
|
---|
15 | % K NxM kernel matrix
|
---|
16 | % KK NxMxK matrix of intermediate kernels
|
---|
17 | %
|
---|
18 | % DESCRIPTION
|
---|
19 | % Derives a string kernel matrix K following the idea of Lodhi et al [2002]. A and B are |
---|
20 | % either individual strings or cell structures of strings. The basic idea is to compute |
---|
21 | % the kernel as an inner product of the feature vectors for two strings and give a sum |
---|
22 | % over all common subsequences weighted according to their frequency of occurrence and |
---|
23 | % lengths. Dynamic programming is used to compute the kernel values between strings. |
---|
24 | % If dataset is needed, it should be created afterwards.
|
---|
25 | %
|
---|
26 | % DEFAULT
|
---|
27 | % B = A
|
---|
28 | % W = [1 1]
|
---|
29 | % LAM = 1
|
---|
30 | % KNORM = 1
|
---|
31 | %
|
---|
32 | % REFERENCE
|
---|
33 | % H.Lodhi, C.Saunders, J. Shawe-Taylor, N.Cristianini, C.Watkins, "Text
|
---|
34 | % Classification using String Kernels", J. of Machine Learning Research 2,
|
---|
35 | % 419-444, 2002.
|
---|
36 | %
|
---|
37 |
|
---|
38 | % Copyright: Elzbieta Pekalska, ela.pekalska@googlemail.com
|
---|
39 | % EWI Faculty, Delft University of Technology and
|
---|
40 | % School of Computer Science, University of Manchester
|
---|
41 |
|
---|
42 |
|
---|
43 | function [K,KK] = strkerm (a,b,w,lam,Knorm)
|
---|
44 | if nargin < 5,
|
---|
45 | Knorm = 1;
|
---|
46 | end
|
---|
47 | if nargin < 4 | isempty(lam),
|
---|
48 | lam = 1;
|
---|
49 | end
|
---|
50 | if nargin < 3 | isempty(w),
|
---|
51 | w = [1 1];
|
---|
52 | end
|
---|
53 | if nargin < 2 | isempty(b),
|
---|
54 | b = a;
|
---|
55 | end
|
---|
56 |
|
---|
57 | if any(w) < 0,
|
---|
58 | error('Weights should be nonnegative.');
|
---|
59 | end
|
---|
60 |
|
---|
61 | if ~iscell(a),
|
---|
62 | if isstr(a) | (~isstr(a) & min(size(a)) == 1)
|
---|
63 | a = {a};
|
---|
64 | else
|
---|
65 | error('A is improper.');
|
---|
66 | end
|
---|
67 | end
|
---|
68 | if ~iscell(b),
|
---|
69 | if isstr(b) | (~isstr(b) & min(size(b)) == 1)
|
---|
70 | b = {b};
|
---|
71 | else
|
---|
72 | error('B is improper.');
|
---|
73 | end
|
---|
74 | end
|
---|
75 |
|
---|
76 |
|
---|
77 | m = length(a);
|
---|
78 | n = length(b);
|
---|
79 |
|
---|
80 |
|
---|
81 | K = zeros(m,n);
|
---|
82 | for i=1:m
|
---|
83 | for j=1:n
|
---|
84 | K(i,j) = strkernel(a{i},b{j},w,lam);
|
---|
85 | end
|
---|
86 | end
|
---|
87 |
|
---|
88 | if Knorm,
|
---|
89 | for i=1:m
|
---|
90 | Kaa(i,1) = strkernel(a{i},a{i},w,lam);
|
---|
91 | end
|
---|
92 | for j=1:n
|
---|
93 | Kbb(1,j) = strkernel(b{j},b{j},w,lam);
|
---|
94 | end
|
---|
95 | K = K ./ sqrt(Kaa*Kbb);
|
---|
96 | end
|
---|
97 | return;
|
---|
98 |
|
---|
99 |
|
---|
100 |
|
---|
101 |
|
---|
102 |
|
---|
103 |
|
---|
104 | function K = strkernel(s,t,w,lam)
|
---|
105 |
|
---|
106 | if isstr(s),
|
---|
107 | s = ['a',s];
|
---|
108 | t = ['a',t];
|
---|
109 | else
|
---|
110 | if size(s,1) > size(s,2)
|
---|
111 | s = [s(1); s];
|
---|
112 | t = [s(1); t];
|
---|
113 | else
|
---|
114 | s = [s(1) s];
|
---|
115 | t = [s(1) t];
|
---|
116 | end
|
---|
117 | end
|
---|
118 |
|
---|
119 | ns = length(s);
|
---|
120 | nt = length(t);
|
---|
121 | n = length(w);
|
---|
122 |
|
---|
123 | K = 0;
|
---|
124 | KK(:,:,1) = ones(ns+1,nt+1);
|
---|
125 | for i=1:n
|
---|
126 | KK(:,:,i+11) = zeros(ns+1,nt+1);
|
---|
127 | for j=1:ns
|
---|
128 | ss = 0;
|
---|
129 | for k=1:nt
|
---|
130 | if t(k) == s(j),
|
---|
131 | ss = ss + KK(j,k,i);
|
---|
132 | end
|
---|
133 | KK(j+1,k+1,i+1) = KK(j,k+1,i+1) +ss;
|
---|
134 | end
|
---|
135 | end
|
---|
136 | K = K + w(i)*KK(ns+1,nt+1,i+1)*lam^(2*i);
|
---|
137 | end
|
---|