rmontanana/BayesNet
rmontanana/BayesNet
1
0
Fork 0
Code Issues 2 Pull Requests Packages Projects Releases 3 Wiki Activity

Remove duplicated code in BestResults

Browse Source
This commit is contained in:
Ricardo Montañana Gómez 2023-09-28 00:59:34 +02:00
parent ac89a451e3
commit 3b06534327
Signed by: rmontanana
GPG Key ID: 46064262FD9A7ADE
4 changed files with 25 additions and 236 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

246
src/Platform/BestResults.cc
View File

@ -7,8 +7,6 @@
#include "Result.h"
#include "Colors.h"
#include "Statistics.h"
#include <boost/math/distributions/chi_squared.hpp>
#include <boost/math/distributions/normal.hpp>
@ -27,12 +25,6 @@ std::string ftime_to_string(TP tp)
buffer << std::put_time(gmt, "%Y-%m-%d %H:%M");
return buffer.str();
}
struct WTL {
int win;
int tie;
int loss;
};
namespace platform {
string BestResults::build()
@ -114,9 +106,10 @@ namespace platform {
}
throw invalid_argument("Unable to open result file. [" + fileName + "]");
}
set<string> BestResults::getModels()
vector<string> BestResults::getModels()
{
set<string> models;
vector<string> result;
auto files = loadResultFiles();
if (files.size() == 0) {
cerr << Colors::MAGENTA() << "No result files were found!" << Colors::RESET() << endl;
@ -129,7 +122,8 @@ namespace platform {
// add the model to the vector of models
models.insert(fileModel);
}
return models;
result = vector<string>(models.begin(), models.end());
return result;
}
void BestResults::buildAll()
@ -171,7 +165,7 @@ namespace platform {
odd = !odd;
}
}
json BestResults::buildTableResults(set<string> models)
json BestResults::buildTableResults(vector<string> models)
{
int numberOfDatasets = 0;
bool first = true;
@ -208,168 +202,8 @@ namespace platform {
table["dateTable"] = ftime_to_string(maxDate);
return table;
}
map<string, float> assignRanks(vector<pair<string, double>>& ranksOrder)
{
// sort the ranksOrder vector by value
sort(ranksOrder.begin(), ranksOrder.end(), [](const pair<string, double>& a, const pair<string, double>& b) {
return a.second > b.second;
});
//Assign ranks to values and if they are the same they share the same averaged rank
map<string, float> ranks;
for (int i = 0; i < ranksOrder.size(); i++) {
ranks[ranksOrder[i].first] = i + 1.0;
}
int i = 0;
while (i < static_cast<int>(ranksOrder.size())) {
int j = i + 1;
int sumRanks = ranks[ranksOrder[i].first];
while (j < static_cast<int>(ranksOrder.size()) && ranksOrder[i].second == ranksOrder[j].second) {
sumRanks += ranks[ranksOrder[j++].first];
}
if (j > i + 1) {
float averageRank = (float)sumRanks / (j - i);
for (int k = i; k < j; k++) {
ranks[ranksOrder[k].first] = averageRank;
}
}
i = j;
}
return ranks;
}
map<int, WTL> computeWTL(int controlIdx, vector<string> models, json table)
{
// Compute the WTL matrix
map<int, WTL> wtl;
int nModels = models.size();
for (int i = 0; i < nModels; ++i) {
wtl[i] = { 0, 0, 0 };
}
json origin = table.begin().value();
for (auto const& item : origin.items()) {
auto controlModel = models.at(controlIdx);
double controlValue = table[controlModel].at(item.key()).at(0).get<double>();
for (int i = 0; i < nModels; ++i) {
if (i == controlIdx) {
continue;
}
double value = table[models[i]].at(item.key()).at(0).get<double>();
if (value < controlValue) {
wtl[i].win++;
} else if (value == controlValue) {
wtl[i].tie++;
} else {
wtl[i].loss++;
}
}
}
return wtl;
}
void postHocHolm(int controlIdx, vector<string> models, int nDatasets, map<string, float> ranks, double significance, map<int, WTL> wtl)
{
// Reference https://link.springer.com/article/10.1007/s44196-022-00083-8
// Post-hoc Holm test
// Calculate the p-value for the models paired with the control model
int nModels = models.size();
map<int, double> stats; // p-value of each model paired with the control model
boost::math::normal dist(0.0, 1.0);
double diff = sqrt(nModels * (nModels + 1) / (6.0 * nDatasets));
for (int i = 0; i < nModels; i++) {
if (i == controlIdx) {
stats[i] = 0.0;
continue;
}
double z = abs(ranks.at(models[controlIdx]) - ranks.at(models[i])) / diff;
double p_value = (long double)2 * (1 - cdf(dist, z));
stats[i] = p_value;
}
// Sort the models by p-value
vector<pair<int, double>> statsOrder;
for (const auto& stat : stats) {
statsOrder.push_back({ stat.first, stat.second });
}
sort(statsOrder.begin(), statsOrder.end(), [](const pair<int, double>& a, const pair<int, double>& b) {
return a.second < b.second;
});
// Holm adjustment
for (int i = 0; i < statsOrder.size(); ++i) {
auto item = statsOrder.at(i);
double before = i == 0 ? 0.0 : statsOrder.at(i - 1).second;
double p_value = min((double)1.0, item.second * (nModels - i));
p_value = max(before, p_value);
statsOrder[i] = { item.first, p_value };
}
cout << Colors::CYAN();
cout << " *************************************************************************************************************" << endl;
cout << " Post-hoc Holm test: H0: 'There is no significant differences between the control model and the other models.'" << endl;
cout << " Control model: " << models[controlIdx] << endl;
cout << " Model p-value rank win tie loss" << endl;
cout << " ============ ============ ========= === === ====" << endl;
// sort ranks from lowest to highest
vector<pair<string, float>> ranksOrder;
for (const auto& rank : ranks) {
ranksOrder.push_back({ rank.first, rank.second });
}
sort(ranksOrder.begin(), ranksOrder.end(), [](const pair<string, float>& a, const pair<string, float>& b) {
return a.second < b.second;
});
for (const auto& item : ranksOrder) {
if (item.first == models.at(controlIdx)) {
continue;
}
auto idx = distance(models.begin(), find(models.begin(), models.end(), item.first));
double pvalue = 0.0;
for (const auto& stat : statsOrder) {
if (stat.first == idx) {
pvalue = stat.second;
}
}
cout << " " << left << setw(12) << item.first << " " << setprecision(10) << fixed << pvalue << setprecision(7) << " " << item.second;
cout << " " << right << setw(3) << wtl.at(idx).win << " " << setw(3) << wtl.at(idx).tie << " " << setw(4) << wtl.at(idx).loss << endl;
}
cout << " *************************************************************************************************************" << endl;
cout << Colors::RESET();
}
bool friedmanTest(vector<string> models, int nDatasets, map<string, float> ranks, double significance = 0.05)
{
// Friedman test
// Calculate the Friedman statistic
int nModels = models.size();
if (nModels < 3 || nDatasets < 3) {
throw runtime_error("Can't make the Friedman test with less than 3 models and/or less than 3 datasets.");
}
cout << Colors::BLUE() << endl;
cout << "***************************************************************************************************************" << endl;
cout << Colors::GREEN() << "Friedman test: H0: 'There is no significant differences between all the classifiers.'" << Colors::BLUE() << endl;
double degreesOfFreedom = nModels - 1.0;
double sumSquared = 0;
for (const auto& rank : ranks) {
sumSquared += pow(rank.second, 2);
}
// Compute the Friedman statistic as in https://link.springer.com/article/10.1007/s44196-022-00083-8
double friedmanQ = 12.0 * nDatasets / (nModels * (nModels + 1)) * (sumSquared - (nModels * pow(nModels + 1, 2)) / 4);
cout << "Friedman statistic: " << friedmanQ << endl;
// Calculate the critical value
boost::math::chi_squared chiSquared(degreesOfFreedom);
long double p_value = (long double)1.0 - cdf(chiSquared, friedmanQ);
double criticalValue = quantile(chiSquared, 1 - significance);
std::cout << "Critical Chi-Square Value for df=" << fixed << (int)degreesOfFreedom
<< " and alpha=" << setprecision(2) << fixed << significance << ": " << setprecision(7) << scientific << criticalValue << std::endl;
cout << "p-value: " << scientific << p_value << " is " << (p_value < significance ? "less" : "greater") << " than " << setprecision(2) << fixed << significance << endl;
bool result;
if (p_value < significance) {
cout << Colors::GREEN() << "The null hypothesis H0 is rejected." << endl;
result = true;
} else {
cout << Colors::YELLOW() << "The null hypothesis H0 is accepted. Computed p-values will not be significant." << endl;
result = false;
}
cout << Colors::BLUE() << "***************************************************************************************************************" << endl;
return result;
}
void BestResults::printTableResults(set<string> models, json table)
void BestResults::printTableResults(vector<string> models, json table)
{
cout << Colors::GREEN() << "Best results for " << score << " as of " << table.at("dateTable").get<string>() << endl;
cout << "------------------------------------------------" << endl;
@ -386,8 +220,6 @@ namespace platform {
auto i = 0;
bool odd = true;
map<string, double> totals;
map<string, float> ranks;
map<string, float> ranksTotal;
int nDatasets = table.begin().value().size();
for (const auto& model : models) {
totals[model] = 0.0;
@ -398,23 +230,12 @@ namespace platform {
cout << color << setw(3) << fixed << right << i++ << " ";
cout << setw(25) << left << item.key() << " ";
double maxValue = 0;
vector<pair<string, double>> ranksOrder;
// Find out the max value for this dataset
for (const auto& model : models) {
double value = table[model].at(item.key()).at(0).get<double>();
if (value > maxValue) {
maxValue = value;
}
ranksOrder.push_back({ model, value });
}
// Assign the ranks
ranks = assignRanks(ranksOrder);
if (ranksTotal.size() == 0) {
ranksTotal = ranks;
} else {
for (const auto& rank : ranks) {
ranksTotal[rank.first] += rank.second;
}
}
// Print the row with red colors on max values
for (const auto& model : models) {
@ -425,7 +246,6 @@ namespace platform {
}
totals[model] += value;
cout << efectiveColor << setw(12) << setprecision(10) << fixed << value << " ";
// cout << efectiveColor << setw(12) << setprecision(10) << fixed << ranks[model] << " ";
}
cout << endl;
odd = !odd;
@ -449,50 +269,7 @@ namespace platform {
}
cout << efectiveColor << setw(12) << setprecision(9) << fixed << totals[model] << " ";
}
// Output the averaged ranks
cout << endl;
int min = 1;
for (auto& rank : ranksTotal) {
if (rank.second < min) {
min = rank.second;
}
rank.second /= nDatasets;
}
cout << Colors::BLUE() << setw(30) << " Ranks....................";
for (const auto& model : models) {
string efectiveColor = Colors::BLUE();
if (ranksTotal[model] == min) {
efectiveColor = Colors::RED();
}
cout << efectiveColor << setw(12) << setprecision(4) << fixed << (double)ranksTotal[model] << " ";
}
cout << endl;
cout << Colors::GREEN() << setw(30) << " Averaged ranks...........";
for (const auto& model : models) {
string efectiveColor = Colors::GREEN();
if (ranksTotal[model] == min) {
efectiveColor = Colors::RED();
}
cout << efectiveColor << setw(12) << setprecision(9) << fixed << (double)ranksTotal[model] << " ";
}
cout << endl;
vector<string> vModels(models.begin(), models.end());
vector<string> datasets;
for (const auto& dataset : table.begin().value().items()) {
datasets.push_back(dataset.key());
}
double significance = 0.05;
if (friedman) {
friedmanTest(vModels, nDatasets, ranksTotal, significance);
// Stablish the control model as the one with the lowest averaged rank
int controlIdx = distance(ranks.begin(), min_element(ranks.begin(), ranks.end(), [](const auto& l, const auto& r) { return l.second < r.second; }));
auto wtl = computeWTL(controlIdx, vModels, table);
postHocHolm(controlIdx, vModels, nDatasets, ranksTotal, significance, wtl);
}
Statistics stats(vModels, datasets, table, significance);
stats.friedmanTest();
stats.postHocHolmTest();
}
void BestResults::reportAll()
{
@ -501,5 +278,16 @@ namespace platform {
json table = buildTableResults(models);
// Print the table of results
printTableResults(models, table);
// Compute the Friedman test
if (friedman) {
vector<string> datasets;
for (const auto& dataset : table.begin().value().items()) {
datasets.push_back(dataset.key());
}
double significance = 0.05;
Statistics stats(models, datasets, table, significance);
auto result = stats.friedmanTest();
stats.postHocHolmTest(result);
}
}
}

6
src/Platform/BestResults.h
View File

@ -14,10 +14,10 @@ namespace platform {
void reportAll();
void buildAll();
private:
set<string> getModels();
vector<string> getModels();
vector<string> loadResultFiles();
json buildTableResults(set<string> models);
void printTableResults(set<string> models, json table);
json buildTableResults(vector<string> models);
void printTableResults(vector<string> models, json table);
string bestResultFile();
json loadFile(const string& fileName);
string path;

7
src/Platform/Statistics.cc
View File

@ -102,7 +102,7 @@ namespace platform {
}
}
void Statistics::postHocHolmTest()
void Statistics::postHocHolmTest(bool friedmanResult)
{
if (!fitted) {
fit();
@ -139,7 +139,8 @@ namespace platform {
p_value = max(before, p_value);
statsOrder[i] = { item.first, p_value };
}
cout << Colors::MAGENTA();
auto color = friedmanResult ? Colors::GREEN() : Colors::YELLOW();
cout << color;
cout << " *************************************************************************************************************" << endl;
cout << " Post-hoc Holm test: H0: 'There is no significant differences between the control model and the other models.'" << endl;
cout << " Control model: " << models[controlIdx] << endl;
@ -203,7 +204,7 @@ namespace platform {
cout << Colors::YELLOW() << "The null hypothesis H0 is accepted. Computed p-values will not be significant." << endl;
result = false;
}
cout << Colors::BLUE() << "***************************************************************************************************************" << endl;
cout << Colors::BLUE() << "***************************************************************************************************************" << Colors::RESET() << endl;
return result;
}
} // namespace platform

2
src/Platform/Statistics.h
View File

@ -17,7 +17,7 @@ namespace platform {
public:
Statistics(vector<string>& models, vector<string>& datasets, json data, double significance = 0.05);
bool friedmanTest();
void postHocHolmTest();
void postHocHolmTest(bool friedmanResult);
private:
void fit();
void computeRanks();