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
|
#include <BALL/QSAR/automaticModelCreator.h>
#include <BALL/QSAR/registry.h>
#include <BALL/QSAR/featureSelection.h>
using namespace BALL::QSAR;
using namespace std;
AutomaticModelCreator::AutomaticModelCreator(const QSARData* data)
{
data_ = data;
min_quality_ = 0.45;
}
void AutomaticModelCreator::setData(const QSARData* data)
{
data_ = data;
}
void AutomaticModelCreator::setMinQuality(double min_quality)
{
min_quality_ = min_quality;
}
void AutomaticModelCreator::optimizeParameters(Model* model)
{
try
{
model->optimizeParameters(5, 20);
KernelModel* km = dynamic_cast<KernelModel*>(model);
if (km) km->kernel->gridSearch(0.25, 20, 0, 5);
}
catch(Exception::SingularMatrixError e)
{
// do nothing; we just cannot optimize the model before feature selection
}
}
void AutomaticModelCreator::selectFeatures(Model* model)
{
FeatureSelection fs(*model);
double threshold = 0.95;
fs.removeHighlyCorrelatedFeatures(threshold);
optimizeParameters(model);
fs.twinScan(5);
optimizeParameters(model);
fs.forwardSelection(5);
optimizeParameters(model);
}
Model* AutomaticModelCreator::generateModel()
{
Registry registry;
bool use_regression = !data_->checkforDiscreteY();
bool use_random_testsets = 0;
Size no_folds = 3;
if (use_random_testsets) no_folds = 10;
Size best_model_id = 0;
Size best_kernel_id = 0;
double best_nested_quality = 0;
/// Do 4-fold nested cross validation, including feature selection steps, for each model type and, if applicable, each kernel function
Log.level(10)<<"model-name kernel #features nested Q^2 stddev"<<std::endl;
Log.level(10)<<"---------------------------------------------------"<<std::endl;
Log.level(10)<<std::setiosflags(std::ios::fixed)<<std::left;
for (Size model_id = 1; model_id < 14; model_id++)
{
RegistryEntry* reg_entry;
try
{
reg_entry = registry.getEntry(model_id);
}
catch(BALL::Exception::GeneralException e)
{
// a model with the current id does not exist
continue;
}
if (use_regression != reg_entry->regression) continue;
Size no_kernel_types = 1;
if (reg_entry->kernel) no_kernel_types = 3;
for (Size kernel_id = 1; kernel_id <= no_kernel_types; kernel_id++)
{
double nested_q2 = 0;
int no_features = 0;
vector<double> q2_values;
for (Size fold_id = 0; fold_id < no_folds; fold_id++)
{
vector<QSARData*> sets;
if (use_random_testsets)
{
// randomly select 25% of compounds for external validation set
sets = data_->generateExternalSet(0.25);
}
else
{
sets = data_->evenSplit(no_folds, fold_id);
}
if (data_->isDataCentered())
{
bool center_y = data_->isResponseCentered();
sets[0]->centerData(center_y); // train-partition
sets[1]->centerData(center_y); // test-partition
}
Model* model;
if (!reg_entry->kernel) model = (*reg_entry->create)(*sets[0]);
else model = (*reg_entry->createKernel1)(*sets[0], kernel_id, 1, -1);
model->setParameters(reg_entry->parameterDefaults);
optimizeParameters(model);
// select relevant features using training partition
selectFeatures(model);
no_features += model->getDescriptorIDs()->size();
// train model using only the training partition
model->readTrainingData();
model->train();
// test fit to external validation data
model->data = sets[1];
model->model_val->testInputData(true);
nested_q2 += model->model_val->getFitRes();
q2_values.push_back(model->model_val->getFitRes());
delete sets[0];
delete sets[1];
delete model;
}
nested_q2 /= no_folds;
no_features /= no_folds;
double stddev = Statistics::getStddev(q2_values);
Log.level(10)<<std::setw(10)<<reg_entry->name_abreviation<<" ";
if (reg_entry->kernel)
{
if (kernel_id == 1) Log.level(10)<<setw(6)<<"polyn.";
else if (kernel_id == 2) Log.level(10)<<setw(6)<<"rbf";
else if (kernel_id == 3) Log.level(10)<<setw(6)<<"sigm.";
}
else Log.level(10)<<setw(6)<<"none";
Log.level(10)<<" ";
Log.level(10)<<setw(9)<<no_features<<" "<<setw(10)<<nested_q2<<" "<<setw(6)<<stddev<<endl<<flush;
double quality = nested_q2-stddev; // make sure to prefer models w/ low stddev
if (quality > best_nested_quality)
{
best_nested_quality = quality;
best_model_id = model_id;
best_kernel_id = kernel_id;
}
}
}
if (best_nested_quality < min_quality_)
{
Log.level(10)<<"Sorry, no model with satisfactory prediction quality found!"<<endl;
return 0;
}
/// Create best model using ENTIRE data set and return it
RegistryEntry* reg_entry = registry.getEntry(best_model_id);
Model* model;
if (!reg_entry->kernel) model = (*reg_entry->create)(*data_);
else model = (*reg_entry->createKernel1)(*data_, best_kernel_id, 1, -1);
model->setParameters(reg_entry->parameterDefaults);
optimizeParameters(model);
selectFeatures(model);
model->readTrainingData();
model->train();
return model;
}
|
