-
Notifications
You must be signed in to change notification settings - Fork 0
/
search_params_new.m
252 lines (224 loc) · 10.2 KB
/
search_params_new.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
clear;
clc;
data_sets = {'flags','Image','yeast','Arts','business','cal500','enron','genbase','llog','medical','slashdot','society'};
% 记录已发表的评估结果作为参考
best_published_result = cell(length(data_sets));
% 评估指标：'HammingLoss', 'RankingLoss', 'OneError', 'Coverage', 'AveragePrecision'
best_published_result{1} = [0.264, 0.227, 0.243, 0.543, 0.804]; % flag X
best_published_result{2} = [0.174, 0.161, 0.306, 0.182, 0.802]; % image Okay
best_published_result{3} = [0.203, 0.172, 0.230, 0.456, 0.763]; % yeast Okay
best_published_result{4} = [0.053, 0.104, 0.444, 0.161, 0.642]; % Arts X
best_published_result{5} = [0.024, 0.030, 0.102, 0.063, 0.899]; % business fix
best_published_result{6} = [0.138, 0.178, 0.117, 0.744, 0.506]; % cal500 Okay
best_published_result{7} = [0.046, 0.070, 0.212, 0.208, 0.715]; % enron X
best_published_result{8} = [0.001, 0.003, 0.000, 0.015, 0.996]; % genbase Okay
best_published_result{9} = [0.015, 0.120, 0.669, 0.133, 0.426]; % llog Okay
best_published_result{10} = [0.010, 0.016, 0.131, 0.027, 0.904]; % medical X
best_published_result{11} = [0.016, 0.042, 0.087, 0.039, 0.897]; % slashdot Okay
best_published_result{12} = [0.051, 0.112, 0.391, 0.178, 0.650]; % society fix
dataset_range = [2,3,5];
for experiment = dataset_range % 选择数据集
BestResult = [];
BestParameter = [];
%% 读取数据集，对数据预处理
dataset_name = data_sets{experiment};
try
cd('datasets');
eval(['load ', dataset_name]);
eval(['load ', dataset_name, '_processed']);
cd('..');
catch exception
fprintf("加载数据文件出错，可能是文件路径问题！\n");
return;
end
if experiment < 4
features = zscore(features);
end
[num_instance, num_feature] = size(features);
lastcol = ones(num_instance, 1);
features = [features lastcol]; %#ok<*AGROW>
%% 参数搜索范围
lambda1_range = [0.0001,0.0003,0.001,0.003,0.1,0.3]; % regularization item
lambda2_range = 10.^(1:-1:-5); % global correlation
lambda3_range = 10.^(3:-1:-3); % local correlation
opt_params.lambda1 = 0;
opt_params.lambda2 = 0;
opt_params.lambda3 = 0;
opt_params.m = 30;
%% 参数评估
BestLambda1=0.003;
BestLambda2=0;
BestLambda3=0;
index = 0;
total = length(lambda1_range)+length(lambda2_range)+length(lambda3_range)+18*2;
%% 搜索最优局部相关系数lambda3
BestResult_temp=[];
for i=1:length(lambda3_range)
index = index + 1;
opt_params.lambda1 = BestLambda1;
opt_params.lambda2 = BestLambda2;
opt_params.lambda3 = lambda3_range(i);
fprintf('%dth / %d %s --search params, lambda1 = %f, lambda2 = %f, lambda3 = %f, m = %f %s \n',index, total, dataset_name, opt_params.lambda1, opt_params.lambda2, opt_params.lambda3, opt_params.m, datestr(now));
Result = egllc_evaluation(opt_params,features,labels,tenfold);
fprintf('Result: %.4f, %.4f, %.4f, %.4f, %.4f \n',Result(1,1), Result(1,2), Result(1,3), Result(1,4), Result(1,5));
% 和临时最优结果比较
if isempty(BestResult_temp) || IsBetterThanBefore(BestResult_temp,Result)==1
BestLambda3 = opt_params.lambda3;
BestResult_temp = Result;
tryTimes=0;
else
tryTimes=tryTimes+1;
if tryTimes == 1
for j = [90:-10:20,9:-1:2]
index = index + 1;
opt_params.lambda3 = j * lambda3_range(i);
fprintf('%dth / %d %s --search params, lambda1 = %f, lambda2 = %f, lambda3 = %f, m = %f %s \n',index, total, dataset_name, opt_params.lambda1, opt_params.lambda2, opt_params.lambda3, opt_params.m, datestr(now));
Result = egllc_evaluation(opt_params,features,labels,tenfold);
fprintf('Result: %.4f, %.4f, %.4f, %.4f, %.4f \n',Result(1,1), Result(1,2), Result(1,3), Result(1,4), Result(1,5));
if IsBetterThanBefore(BestResult_temp,Result)==1
BestLambda3 = opt_params.lambda3;
BestResult_temp = Result;
end
end
end
end
end
%% 搜索最优全局相关系数lambda2
BestResult_temp=[];
for i=1:length(lambda2_range)
index = index + 1;
opt_params.lambda1 = BestLambda1;
opt_params.lambda2 = lambda2_range(i);
opt_params.lambda3 = BestLambda3;
fprintf('%dth / %d %s --search params, lambda1 = %f, lambda2 = %f, lambda3 = %f, m = %f %s \n',index, total, dataset_name, opt_params.lambda1, opt_params.lambda2, opt_params.lambda3, opt_params.m, datestr(now));
Result = egllc_evaluation(opt_params,features,labels,tenfold);
fprintf('Result: %.4f, %.4f, %.4f, %.4f, %.4f \n',Result(1,1), Result(1,2), Result(1,3), Result(1,4), Result(1,5));
% 和临时最优结果比较
if isempty(BestResult_temp) || IsBetterThanBefore(BestResult_temp,Result)==1
BestLambda2 = opt_params.lambda2;
BestResult_temp = Result;
tryTimes=0;
else
tryTimes=tryTimes+1;
if tryTimes == 1
for j = [90:-10:20,9:-1:2]
index = index + 1;
opt_params.lambda2 = j * lambda2_range(i);
fprintf('%dth / %d %s --search params, lambda1 = %f, lambda2 = %f, lambda3 = %f, m = %f %s \n',index, total, dataset_name, opt_params.lambda1, opt_params.lambda2, opt_params.lambda3, opt_params.m, datestr(now));
Result = egllc_evaluation(opt_params,features,labels,tenfold);
fprintf('Result: %.4f, %.4f, %.4f, %.4f, %.4f \n',Result(1,1), Result(1,2), Result(1,3), Result(1,4), Result(1,5));
if IsBetterThanBefore(BestResult_temp,Result)==1
BestLambda2 = opt_params.lambda2;
BestResult_temp = Result;
end
end
end
end
end
%% 搜索最优正则化项
BestResult_temp=[];
for i=1:length(lambda1_range)
index = index + 1;
opt_params.lambda1 = lambda1_range(i);
opt_params.lambda2 = BestLambda2;
opt_params.lambda3 = BestLambda3;
fprintf('%dth / %d %s --search params, lambda1 = %f, lambda2 = %f, lambda3 = %f, m = %f %s \n',index, total, dataset_name, opt_params.lambda1, opt_params.lambda2, opt_params.lambda3, opt_params.m, datestr(now));
Result = egllc_evaluation(opt_params,features,labels,tenfold);
fprintf('Result: %.4f, %.4f, %.4f, %.4f, %.4f \n',Result(1,1), Result(1,2), Result(1,3), Result(1,4), Result(1,5));
% 和临时最优结果比较
if isempty(BestResult_temp) || IsBetterThanBefore(BestResult_temp,Result)==1
BestLamba1 = opt_params.lambda1;
BestResult_temp = Result;
end
end
BestResult = BestResult_temp;
BestParameter.lambda1 = BestLambda1;
BestParameter.lambda2 = BestLambda2;
BestParameter.lambda3 = BestLambda3;
BestParameter.m = opt_params.m;
%% 存储该数据集最优参数
% 和已发表结果比较
r = IsBetterThanBefore(best_published_result{experiment},BestResult);
if r == 1
isFind = 1;
fprintf('Find parameters better than before!\n');
else
isFind = 0;
fprintf('Can not find parameters better than before!\n');
end
fprintf('Best params, lambda1 = %f, lambda2 = %f, lambda3 = %f, gamma = %f \n', BestParameter.lambda1, BestParameter.lambda2, BestParameter.lambda3, BestParameter.m);
fprintf('BestResult: %.4f, %.4f, %.4f, %.4f, %.4f \n\n',BestResult(1,1), BestResult(1,2), BestResult(1,3), BestResult(1,4), BestResult(1,5));
cd('parameters_search');
eval(['save ', dataset_name, '_parameters.mat BestParameter BestResult isFind']);
cd('..');
% 调用实验函数，验证当前数据集
logistic_func(experiment)
end
% logistic_func(dataset_range)
function result = egllc_evaluation(parms,features,labels,tenfold)
% 数据集被分为5折，这里使用4折,方便并行
cross = 4;
temp_result = zeros(cross, 5);
selected = randi([1,10],1);
indices = tenfold{selected};
parfor rep=1:cross
testIdx = find(indices == rep);
trainIdx = setdiff(find(indices),testIdx);
test_x = features(testIdx,:);
test_target = labels(testIdx,:); %#ok<*PFBNS>
train_x = features(trainIdx,:);
train_target = labels(trainIdx,:);
% feature_cluster = {};
% label_cluster = {};
[Idx,~] = kmeans(train_x,parms.m);
local_size = zeros(1, parms.m);
all_local_L = cell(1, parms.m);
for cluster_id = 1:parms.m
fea = train_x(Idx==cluster_id, :); %#ok<*IJCL>
lab = train_target(Idx==cluster_id, :);
local_size(cluster_id) = size(fea,1);
local_y = lab;
local_y(local_y == 0) = -1;
local_c = 1 - pdist2( local_y'+eps, local_y'+eps, 'cosine');
local_D = sum(local_c, 2);
local_L = diag(local_D) - local_c;
all_local_L{cluster_id} = local_L;
end
y = train_target;
y(y==0) = -1; % 把标记中的0变成-1，这样用余弦相似性计算标记相关性时能计算出负相关
corr = 1 - pdist2( y'+eps, y'+eps, 'cosine' );
% 随机初始化权重矩阵
item=rand(size(train_x,2),size(train_target,2));
% 训练，预测并评估
[weights,~] = egllc_train(@(item)lbfgs_progress(item,train_x,train_target,parms.lambda1,parms.lambda2,parms.lambda3,corr,parms.m,local_size, all_local_L),item);
[~, ~, res_once] = egllc_predict(weights,test_x,test_target);
temp_result(rep, :) = res_once;
end
result(1, :) = mean(temp_result, 1);
end
% 对两组评价指标进行综合比较
function r = IsBetterThanBefore(BestResult,CurrentResult)
a = CurrentResult(1,1) + CurrentResult(1,2) + CurrentResult(1,3) + CurrentResult(1,4) - CurrentResult(1,5);
b = BestResult(1,1) + BestResult(1,2) + BestResult(1,3) + BestResult(1,4) - BestResult(1,5);
count = 0;
if BestResult(1,1) >= CurrentResult(1,1)-0.001
count = count+1;
end
if BestResult(1,2) >= CurrentResult(1,2) - 0.001
count = count+1;
end
if BestResult(1,3) >= CurrentResult(1,3) -0.001
count = count+1;
end
if BestResult(1,4) >= CurrentResult(1,4) - 0.001
count = count+1;
end
if BestResult(1,5) <= CurrentResult(1,5) + 0.001
count = count+1;
end
if count>=4 || a<b
r = 1;
else
r = 0;
end
end