public string decide(TransactionDto transaction, DecisionTreeNode node)
{
if (node.split_index == -1) return node.decision;
foreach (DecisionTreeNode child in node.children)
{
if (transaction.items[node.split_index] == child.split_value) return decide(transaction, child);
}
return "NOT_FOUND"; //this case *SHOULD* never be reached
}
private void train(ref DecisionTreeNode root, List<TransactionDto> transactions, int class_index)
{
//clear children
root.children.Clear();
List<DecisionTreeClass> classes = new List<DecisionTreeClass>();
foreach (TransactionDto t in transactions)
{
string value = t.items[class_index];
Boolean found = false;
foreach (DecisionTreeClass c in classes)
{
if (c.value == value)
{
found = true;
c.increment();
}
}
if (!found)
{
classes.Add(new DecisionTreeClass(value));
}
}
int m = classes.Count;
if (m == 1)
{
root.decision = classes[0].value;
return;
}
//compute infoGain for classifier
int totalTuples = transactions.Count();
double infoGain = 0.0;
for (int i = 0; i < m; i++)
{
infoGain -= ((double)classes[i].count / totalTuples) * Math.Log(((double)classes[i].count / totalTuples), 2);
}
List<double> attributeGain = new List<double>();
//compute infoGain for attributes
for (int i = 0; i < transactions[0].items.Count; i++)
{
if (i == class_index) continue; //skip classifier
var distinctValues = from v in transactions
select v.items[i];
distinctValues = distinctValues.Distinct().ToList();
double ttlAttrInfoGain = 0.0;
foreach (string d in distinctValues)
{
List<int> counts = new List<int>();
foreach (DecisionTreeClass c in classes)
{
var distinctClassValues = from e in transactions
where e.items[class_index] == c.value && e.items[i] == d
select e;
counts.Add(distinctClassValues.Count());
}
double attrInfoGain = 0.0;
int sum = counts.Sum();
foreach (int c in counts)
{
if (c == 0) continue;
attrInfoGain -= ((double)c / sum) * Math.Log(((double)c / sum), 2);
}
attrInfoGain *= (double)sum / totalTuples;
ttlAttrInfoGain += attrInfoGain;
}
attributeGain.Add(ttlAttrInfoGain);
} //attribute loop
var aGain = from a in attributeGain
select infoGain - a;
attributeGain = aGain.ToList();
//pg 339, need to partition based on max(attrGain)
int splitIndex = attributeGain.IndexOf(attributeGain.Max());
var splitValues = from v in transactions
select v.items[splitIndex];
splitValues = splitValues.Distinct().ToList();
foreach (string split in splitValues)
{
var nTrans = from t in transactions
where t.items[splitIndex] == split
select t;
nTrans = nTrans.ToList();
root.split_index = splitIndex;
DecisionTreeNode node = new DecisionTreeNode(-1, split);
root.children.Add(node);
train(ref node, (List<TransactionDto>)nTrans, class_index);
}
}
private void train(ref DecisionTreeNode root, List <TransactionDto> transactions, int class_index)
{
//clear children
root.children.Clear();
List <DecisionTreeClass> classes = new List <DecisionTreeClass>();
foreach (TransactionDto t in transactions)
{
string value = t.items[class_index];
Boolean found = false;
foreach (DecisionTreeClass c in classes)
{
if (c.value == value)
{
found = true;
c.increment();
}
}
if (!found)
{
classes.Add(new DecisionTreeClass(value));
}
}
int m = classes.Count;
if (m == 1)
{
root.decision = classes[0].value;
return;
}
//compute infoGain for classifier
int totalTuples = transactions.Count();
double infoGain = 0.0;
for (int i = 0; i < m; i++)
{
infoGain -= ((double)classes[i].count / totalTuples) * Math.Log(((double)classes[i].count / totalTuples), 2);
}
List <double> attributeGain = new List <double>();
//compute infoGain for attributes
for (int i = 0; i < transactions[0].items.Count; i++)
{
if (i == class_index)
{
continue; //skip classifier
}
var distinctValues = from v in transactions
select v.items[i];
distinctValues = distinctValues.Distinct().ToList();
double ttlAttrInfoGain = 0.0;
foreach (string d in distinctValues)
{
List <int> counts = new List <int>();
foreach (DecisionTreeClass c in classes)
{
var distinctClassValues = from e in transactions
where e.items[class_index] == c.value && e.items[i] == d
select e;
counts.Add(distinctClassValues.Count());
}
double attrInfoGain = 0.0;
int sum = counts.Sum();
foreach (int c in counts)
{
if (c == 0)
{
continue;
}
attrInfoGain -= ((double)c / sum) * Math.Log(((double)c / sum), 2);
}
attrInfoGain *= (double)sum / totalTuples;
ttlAttrInfoGain += attrInfoGain;
}
attributeGain.Add(ttlAttrInfoGain);
} //attribute loop
var aGain = from a in attributeGain
select infoGain - a;
attributeGain = aGain.ToList();
//pg 339, need to partition based on max(attrGain)
int splitIndex = attributeGain.IndexOf(attributeGain.Max());
var splitValues = from v in transactions
select v.items[splitIndex];
splitValues = splitValues.Distinct().ToList();
foreach (string split in splitValues)
{
var nTrans = from t in transactions
where t.items[splitIndex] == split
select t;
nTrans = nTrans.ToList();
root.split_index = splitIndex;
DecisionTreeNode node = new DecisionTreeNode(-1, split);
root.children.Add(node);
train(ref node, (List <TransactionDto>)nTrans, class_index);
}
} //train
public DecisionTree(List<TransactionDto> transactions, int class_index)
{
_root = new DecisionTreeNode(-1, String.Empty);
train(ref _root, transactions, class_index);
}
public DecisionTree(List <TransactionDto> transactions, int class_index)
{
_root = new DecisionTreeNode(-1, String.Empty);
train(ref _root, transactions, class_index);
}