-
Notifications
You must be signed in to change notification settings - Fork 1
/
ml_feature_importance.m
269 lines (243 loc) · 9.82 KB
/
ml_feature_importance.m
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
% LDA classification of ERPs in timewise manner among trials
clear all
%%%%%%%%%%%%%
% Stimuli Locked or Response Locked
StimLocked = 1;
if StimLocked
load('/cubric/collab/ccbrain/data/Scripts/eeg_analysis2/Data/GlobalAveragedDataStim.mat')
timeshift = epoch_length(1);
else
load('/cubric/collab/ccbrain/data/Scripts/eeg_analysis2/Data/GlobalAveragedDataResp.mat')
timeshift = epoch_length(1);
end
%%%%%%%%%%%%%
clear GlobalMean GlobalVar GlobalStd
srate = 250;
timex = (1:size(FullData.Control100.u1000,2))/srate + timeshift;
%% Setup configuration struct for LDA
cfg_LDA = [];
cfg_LDA.classifier = 'lda';
cfg_LDA.param = struct('lambda','auto');
cfg_SVM = mv_get_classifier_param('svm');
cfg_SVM.classifier = 'svm';
cfg_SVM.param.C = 0.1;%[0.01 0.1 0.5];
cfg_LR = [];
cfg_LR.classifier = 'logreg';
cfg_LR.param = struct('lambda','auto' );
cfg_libSVM = mv_get_classifier_param('libsvm');
cfg_libSVM.classifier = 'libsvm';
cfg_libSVM.C = 0.1;%[0.01 0.1 0.5];
cfg_libSVM.kernel_type = 0; % 0 linear, 2 rbf
%%%%%%%%%%%%%%%%%%%%%%%%%%Pick classifier
cfg_clf = cfg_SVM;
%%%%%%%%%%%%%%%% Hyperparams
smoothing_window = 0.02; % sec
timestamps = [0.2, 0.35, 0.5];
%%%%%%%%%%%%%%%
%% Loop with model training among the subjects
for sub_idx = 1 : length(subjects) %sub_idx = 1;
subject_code = ['u' num2str(subjects(sub_idx))];
disp(['----------------------- ' subject_code])
tmpA = permute(FullData.Control100.(subject_code),[3 1 2]);
tmpB = permute(FullData.Equal100.(subject_code),[3 1 2]);
tmpA = addEvery2ndRows(tmpA);
tmpB = addEvery2ndRows(tmpB);
edata.compare100 = double([tmpA; tmpB]);
edata.compare100 = applyToDimension(@(xx) gaussfilt(timex, xx, smoothing_window), edata.compare100, 3);
% hold on; plot(squeeze(edata.compare100(2,4,:))); plot(squeeze(xw(2,4,:))); hold off;
labels.compare100= ones(1, size(edata.compare100,1));
labels.compare100(size(tmpA,1):size(edata.compare100,1)) = 2;
tmpA = permute(FullData.Control80.(subject_code),[3 1 2]);
tmpB = permute(FullData.Equal80.(subject_code),[3 1 2]);
tmpA = addEvery2ndRows(tmpA);
tmpB = addEvery2ndRows(tmpB);
edata.compare80 = double([tmpA; tmpB]);
edata.compare80 = applyToDimension(@(xx) gaussfilt(timex, xx, smoothing_window), edata.compare80, 3);
labels.compare80= ones(1, size(edata.compare80,1));
labels.compare80(size(tmpA,1):size(edata.compare80,1)) = 2;
tmpA = permute(FullData.Control20.(subject_code),[3 1 2]);
tmpB = permute(FullData.Equal20.(subject_code),[3 1 2]);
tmpA = addEvery2ndRows(tmpA);
tmpB = addEvery2ndRows(tmpB);
edata.compare20 = double([tmpA; tmpB]);
edata.compare20 = applyToDimension(@(xx) gaussfilt(timex, xx, smoothing_window), edata.compare20, 3);
labels.compare20= ones(1, size(edata.compare20,1));
labels.compare20(size(tmpA,1):size(edata.compare20,1)) = 2;
tmpA = permute(FullData.Equal80.(subject_code),[3 1 2]);
tmpB = permute(FullData.Equal100.(subject_code),[3 1 2]);
tmpA = addEvery2ndRows(tmpA);
tmpB = addEvery2ndRows(tmpB);
edata.compare80100 = double([tmpA; tmpB]);
edata.compare80100 = applyToDimension(@(xx) gaussfilt(timex, xx, smoothing_window), edata.compare80100, 3);
labels.compare80100 = ones(1, size(edata.compare80100,1));
labels.compare80100(size(tmpA,1):size(edata.compare80100,1)) = 2;
tmpA = permute(FullData.Equal20.(subject_code),[3 1 2]);
tmpB = permute(FullData.Equal100.(subject_code),[3 1 2]);
tmpA = addEvery2ndRows(tmpA);
tmpB = addEvery2ndRows(tmpB);
edata.compare20100 = double([tmpA; tmpB]);
edata.compare20100 = applyToDimension(@(xx) gaussfilt(timex, xx, smoothing_window), edata.compare20100, 3);
labels.compare20100 = ones(1, size(edata.compare20100,1));
labels.compare20100(size(tmpA,1):size(edata.compare20100,1)) = 2;
tmpA = permute(FullData.Equal20.(subject_code),[3 1 2]);
tmpB = permute(FullData.Equal80.(subject_code),[3 1 2]);
tmpA = addEvery2ndRows(tmpA);
tmpB = addEvery2ndRows(tmpB);
edata.compare2080 = double([tmpA; tmpB]);
edata.compare2080 = applyToDimension(@(xx) gaussfilt(timex, xx, smoothing_window), edata.compare2080, 3);
labels.compare2080 = ones(1, size(edata.compare2080,1));
labels.compare2080(size(tmpA,1):size(edata.compare2080,1)) = 2;
tmpA = permute(FullData.pref.(subject_code),[3 1 2]);
tmpB = permute(FullData.notpref.(subject_code),[3 1 2]);
tmpA = addEvery2ndRows(tmpA);
tmpB = addEvery2ndRows(tmpB);
edata.comparePref = double([tmpA; tmpB]);
edata.comparePref = applyToDimension(@(xx) gaussfilt(timex, xx, smoothing_window), edata.comparePref, 3);
labels.comparePref = ones(1, size(edata.comparePref,1));
labels.comparePref(size(tmpA,1):size(edata.comparePref,1)) = 2;
avg_w = zeros(length(timestamps), size(edata.compare100,2));
for tt = 1:length(timestamps)
tx = find(timex >= timestamps(tt));
tt_idx = tx(1);
dd = squeeze(edata.compare100(:,:,tt_idx));
cc = train_svm(cfg_clf, dd,labels.compare100);
activation = transformToInterpretable(dd,cc);
avg_w(tt,:) = activation;%cc.w;
end
avg_w = mean(avg_w,1);
all_weights.compare100(sub_idx, :) = avg_w;
%compare80
avg_w = zeros(length(timestamps), size(edata.compare80,2));
for tt = 1:length(timestamps)
tx = find(timex >= timestamps(tt));
tt_idx = tx(1);
dd = squeeze(edata.compare80(:,:,tt_idx));
cc = train_svm(cfg_clf, dd,labels.compare80);
activation = transformToInterpretable(dd,cc);
avg_w(tt,:) = activation;%cc.w;
end
avg_w = mean(avg_w,1);
all_weights.compare80(sub_idx, :) = avg_w;
%compare20
avg_w = zeros(length(timestamps), size(edata.compare20,2));
for tt = 1:length(timestamps)
tx = find(timex >= timestamps(tt));
tt_idx = tx(1);
dd = squeeze(edata.compare20(:,:,tt_idx));
cc = train_svm(cfg_clf, dd,labels.compare20);
activation = transformToInterpretable(dd,cc);
avg_w(tt,:) = activation;%cc.w;
end
avg_w = mean(avg_w,1);
all_weights.compare20(sub_idx, :) = avg_w;
% compare80100
avg_w = zeros(length(timestamps), size(edata.compare80100,2));
for tt = 1:length(timestamps)
tx = find(timex >= timestamps(tt));
tt_idx = tx(1);
dd = squeeze(edata.compare80100(:,:,tt_idx));
cc = train_svm(cfg_clf, dd,labels.compare80100);
activation = transformToInterpretable(dd,cc);
avg_w(tt,:) = activation;%cc.w;
end
avg_w = mean(avg_w,1);
all_weights.compare80100(sub_idx, :) = avg_w;
% compare20100
avg_w = zeros(length(timestamps), size(edata.compare20100,2));
for tt = 1:length(timestamps)
tx = find(timex >= timestamps(tt));
tt_idx = tx(1);
dd = squeeze(edata.compare20100(:,:,tt_idx));
cc = train_svm(cfg_clf, dd,labels.compare20100);
activation = transformToInterpretable(dd,cc);
avg_w(tt,:) = activation;%cc.w;
end
avg_w = mean(avg_w,1);
all_weights.compare20100(sub_idx, :) = avg_w;
% compare2080
avg_w = zeros(length(timestamps), size(edata.compare2080,2));
for tt = 1:length(timestamps)
tx = find(timex >= timestamps(tt));
tt_idx = tx(1);
dd = squeeze(edata.compare2080(:,:,tt_idx));
cc = train_svm(cfg_clf, dd,labels.compare2080);
% From S. Haufe 2014:
activation = transformToInterpretable(dd,cc);
avg_w(tt,:) = activation;%cc.w;
end
avg_w = mean(avg_w,1);
all_weights.compare2080(sub_idx, :) = avg_w;
% comparePREF
avg_w = zeros(length(timestamps), size(edata.comparePref,2));
for tt = 1:length(timestamps)
tx = find(timex >= timestamps(tt));
tt_idx = tx(1);
dd = squeeze(edata.comparePref(:,:,tt_idx));
cc = train_svm(cfg_clf, dd,labels.comparePref);
% From S. Haufe 2014:
activation = transformToInterpretable(dd,cc);
avg_w(tt,:) = activation;%cc.w;
end
avg_w = mean(avg_w,1);
all_weights.comparePref(sub_idx, :) = avg_w;
end
%save(['Data/ml_feature_importance_lsvm'], 'all_weights', 'timex', 'cfg_clf')
%%
% Plotting the resuts
load('Data/chanlocs');
fields = fieldnames(all_weights);
val.max = [];
val.min = [];
for i = 1:length(fields)
val.max = [val.max max(abs(mean(all_weights.(fields{i}))))];
val.min = [val.min min(abs(mean(all_weights.(fields{i}))))];
end
topoplot(abs(mean(all_weights.comparePref,1)), chanlocs32, 'maplimits', [min(val.min), max(val.max)])
figure;
subplot(131)
topoplot(abs(mean(all_weights.compare100,1)), chanlocs32, 'maplimits', [min(val.min), max(val.max)])
title('Equal 100 vs Control 100')
subplot(132)
topoplot(abs(mean(all_weights.compare80,1)), chanlocs32, 'maplimits', [min(val.min), max(val.max)])
title('Equal 80 vs Control 80')
subplot(133)
topoplot(abs(mean(all_weights.compare80,1)), chanlocs32, 'maplimits', [min(val.min), max(val.max)])
title('Equal 20 vs Control 20')
figure;
subplot(131)
topoplot(abs(mean(all_weights.compare80100,1)), chanlocs32, 'maplimits', [min(val.min), max(val.max)])
title('Equal 80 vs Equal 100')
subplot(132)
topoplot(abs(mean(all_weights.compare20100,1)), chanlocs32, 'maplimits', [min(val.min), max(val.max)])
title('Equal 20 vs Equal 100')
subplot(133)
topoplot(abs(mean(all_weights.compare2080,1)), chanlocs32, 'maplimits', [min(val.min), max(val.max)])
title('Equal 20 vs Equal 80')
%%
fields = fieldnames(all_weights);
val.max = [];
val.min = [];
for i = 1:length(fields)
val.max = [val.max max(mean(all_weights.(fields{i})))];
val.min = [val.min min(mean(all_weights.(fields{i})))];
end
figure;
subplot(131)
topoplot(mean(all_weights.compare100,1), chanlocs32, 'maplimits', [min(val.min), max(val.max)])
title('Equal 100 vs Control 100')
subplot(132)
topoplot(mean(all_weights.compare80,1), chanlocs32, 'maplimits', [min(val.min), max(val.max)])
title('Equal 80 vs Control 80')
subplot(133)
topoplot(mean(all_weights.compare80,1), chanlocs32, 'maplimits', [min(val.min), max(val.max)])
title('Equal 20 vs Control 20')
figure;
subplot(131)
topoplot(mean(all_weights.compare80100,1), chanlocs32, 'maplimits', [min(val.min), max(val.max)])
title('Equal 80 vs Equal 100')
subplot(132)
topoplot(mean(all_weights.compare20100,1), chanlocs32, 'maplimits', [min(val.min), max(val.max)])
title('Equal 20 vs Equal 100')
subplot(133)
topoplot(mean(all_weights.compare2080,1), chanlocs32, 'maplimits', [min(val.min), max(val.max)])
title('Equal 20 vs Equal 80')