public void run_KMeans(string file_path, int k, out double train, out double valid, out double test, out double size, out double rmse)
{
Accuracies_training.Clear();
Accuracies_validation.Clear();
Accuracies_testing.Clear();
trees.Clear();
tree_sizes.Clear();
rmse_list.Clear();
DataLoader d = new DataLoader();
DecisionTree tree = new DecisionTree();
Accuracy a = new Accuracy();
d.get_K_Partitions(file_path, k); //fills d.partitions with k even partitions of the dataset (each contains a header row)
for (int i = 0; i < k; i++) //for each partition configuration
{
Console.WriteLine("Partition " + i + " / " + k + " ---------------------------------------------------------------");
List <string> training_data = new List <string>();
List <string> testing_data = new List <string>();
List <string> validation_data = new List <string>();
training_data.Add(d.title_row); //Add title row to top of training set
for (int j = 0; j < k; j++)
{
if (j != i) //Iteratively keep one partition to be used as the test set
{
for (int z = 0; z < d.partitions[j].Length; z++)
{
training_data.Add(d.partitions[j][z]);
}
}
else
{
for (int z = 0; z < d.partitions[j].Length; z++)
{
testing_data.Add(d.partitions[j][z]);
}
}
}
//Reserve 50% of the training data to be the validation set (move the rows to validation_data)
int s = training_data.Count / 2;
validation_data = training_data.GetRange(training_data.Count - s, s);
training_data.RemoveRange(training_data.Count - s, s);
DataTable x = d.CreateTable(training_data.ToArray()); //input: string[] output: DataTable
List <DataColumn> all_attributes = d.getAllAttributes(x);
Node root = tree.root = tree.RunC4_5(x, all_attributes);
root.isRoot = true; //Set identifier of the root
root.pruneTree(root);
trees.Add(root);
training_data.RemoveAt(0);
List <string> validation_subset = getValidationSubset(validation_data);
//Optimise with respect to the validation set
for (int it = 0; it < 10000; it++)
{
/////////////////////////////////////////////////SELECT OBJECTIVE FUNCTION///////////////////////////////////////////////////////////
//root = root.randomMutateAndRebuild_Accuracy(root); //Objective Function: Maximise Accuracy (regardless of size)
//root = root.randomMutateAndRebuild_RMSE(root); //Objective Function: Minimise RMSE (For regression trees, regardless of size)
//PARETO FRONT
//The below objective function is a pareto front. It minimises the size of the tree while also increasing accuracy (if either remain stable, the change is accepted)
if ((it % 100) == 0)
{
validation_subset = getValidationSubset(validation_data);
} //Randomise validation subset every x iterations
root = root.randomMutateAndRebuild_Size(root, validation_subset.ToArray()); //Objective Function: Minimise size of the tree (number of nodes)
//force a mutation here if the accuracy has not improved in the last 100 iterations, for instance...
}
//Save the accuracies of each partition
Accuracies_training.Add(a.GetAccuracy(root, training_data.ToArray()));
Accuracies_validation.Add(a.GetAccuracy(root, validation_data.ToArray()));
Accuracies_testing.Add(a.GetAccuracy(root, testing_data.ToArray()));
tree_sizes.Add(Convert.ToDouble(root.flattenTree(root).Length));
rmse_list.Add(a.getRMSE(root, testing_data.ToArray()));
x.Clear(); //Clear DataTable so that we can begin the next C4.5 run - on the next partition
}
Console.WriteLine("\n\n");
Console.WriteLine("Final report: ");
double training_total = 0;
foreach (double q in Accuracies_training.Reverse <double>())
{
if (q != 0)
{
training_total += q;
}
else
{
Accuracies_training.Remove(q);
}
}
double average_training_accuracy = training_total / Accuracies_training.Count;
double validation_total = 0;
foreach (double q in Accuracies_validation.Reverse <double>())
{
if (q != 0)
{
validation_total += q;
}
else
{
Accuracies_validation.Remove(q);
}
}
double average_validation_accuracy = validation_total / Accuracies_validation.Count;
double testing_total = 0;
double highest_acc = double.NegativeInfinity;
int highest_acc_index = 0;
for (int t = 0; t < Accuracies_testing.Count; t++)
{
if (Accuracies_testing[t] != 0)
{
testing_total += Accuracies_testing[t];
if (Accuracies_testing[t] > highest_acc)
{
highest_acc = Accuracies_testing[t]; highest_acc_index = t;
}
}
else
{
Accuracies_testing.RemoveAt(t);
}
}
double average_testing_accuracy = testing_total / Accuracies_testing.Count;
double tot = 0;
foreach (double i in tree_sizes)
{
tot += i / 2;
}
double average_size = tot / tree_sizes.Count;
double tot_rmse = 0;
foreach (double r in rmse_list.Reverse <double>())
{
if (r != 0)
{
tot_rmse += r;
}
else
{
rmse_list.Remove(r);
}
}
double average_rmse = tot_rmse / rmse_list.Count;
//Set 'out' variables for collection
train = average_training_accuracy;
valid = average_validation_accuracy;
test = average_testing_accuracy;
size = average_size;
rmse = average_rmse;
Console.WriteLine("Training accuracies:");
foreach (double p in Accuracies_training)
{
Console.WriteLine(p);
}
Console.WriteLine("Validation accuracies:");
foreach (double p in Accuracies_validation)
{
Console.WriteLine(p);
}
Console.WriteLine("Testing accuracies:");
foreach (double p in Accuracies_testing)
{
Console.WriteLine(p);
}
Console.WriteLine("Average training accuracy: " + average_training_accuracy);
Console.WriteLine("Average validation accuracy: " + average_validation_accuracy);
Console.WriteLine("Average testing accuracy: " + average_testing_accuracy);
Console.WriteLine("Average tree size: " + average_size);
Console.WriteLine("Printed tree (highest test accuracy) : " + Accuracies_testing[highest_acc_index]);
//Visualise the tree with the highest test accuracy
DOT_file_generator df = new DOT_file_generator();
df.createDOTfile(trees[highest_acc_index]);
}