/**
* Remove FV based on its threshold, if it is over <b> threshold </b> then it would be removed.
* Let the FVs that has entropy below threshold remains.
*
* @param fv_list
* @param threshold
* @return Remaining FVs
*/
public Map<FV, Collection<FV>> applyThresholdRemoval(
Map<FV, Collection<FV>> fv_list, double threshold) {
Map<FV, Collection<FV>> result = new HashMap<FV, Collection<FV>>();
List<Double> entropies = generateEntropy(fv_list);
result.putAll(fv_list);
// See mean, Q1~Q3 values for entropy threshold
Double[] temp = new Double[entropies.size()];
temp = entropies.toArray(temp);
Double mean = MathHelper.getInstance().calculateAverage(temp);
Double stdev = MathHelper.getInstance().calculateStdev(mean, temp);
Double[] q = MathHelper.getInstance().calculateQuartile(temp);
// System.out.println("Mean: " + mean);
// System.out.println("Stdev: " + stdev);
// System.out.println("Mean + Stdev: " + (mean + stdev));
// System.out.println("Mean - Stdev: " + (mean - stdev));
// System.out.println("Q1: " + q[0]);
// System.out.println("Q2: " + q[1]);
// System.out.println("Q3: " + q[2]);
// threshold = mean + stdev; // Force using specified threshold
switch (thr_alg) {
case Mean:
threshold = mean;
break;
case MeanMin:
threshold = mean - stdev;
break;
case MeanPlus:
threshold = mean + stdev;
break;
case Q1:
threshold = q[0];
break;
case Q2:
threshold = q[1];
break;
case Q3:
threshold = q[2];
break;
default:
break;
}
// Apply removal
for (FV k : fv_list.keySet()) {
if (k.getEntropy() > threshold) result.remove(k);
}
return result;
}
// Instead of top-k, we select top-k percents data
public Map<FV, Collection<FV>> applyCorrelationRemoval(
Map<FV, Collection<FV>> fv_list, double topk, Double[][] CM, List<Double> corrValues) {
Map<FV, Collection<FV>> result = new HashMap<FV, Collection<FV>>();
generateEntropy(fv_list);
Double mean;
Double stdev;
Double[] q = new Double[3]; // Q1, Q2, Q3
Double[] temp = new Double[corrValues.size()];
temp = corrValues.toArray(temp);
// Calculate mean, Q1~Q3
mean = MathHelper.getInstance().calculateAverage(temp);
stdev = MathHelper.getInstance().calculateStdev(mean, temp);
q = MathHelper.getInstance().calculateQuartile(temp);
// System.out.println("Mean: " + mean);
// System.out.println("Stdev: " + stdev);
// System.out.println("Mean + Stdev: " + (mean + stdev));
// System.out.println("Mean - Stdev: " + (mean - stdev));
// System.out.println("Q1: " + q[0]);
// System.out.println("Q2: " + q[1]);
// System.out.println("Q3: " + q[2]);
// Selecting correlation threshold
double corrThreshold = q[1]; // Use Q3 / Q2 / Mean+Stdev as corr
// threshold
switch (thr_alg) {
case Mean:
corrThreshold = mean;
break;
case MeanMin:
corrThreshold = mean - stdev;
break;
case MeanPlus:
corrThreshold = mean + stdev;
break;
case Q1:
corrThreshold = q[0];
break;
case Q2:
corrThreshold = q[1];
break;
case Q3:
corrThreshold = q[2];
break;
default:
break;
}
/** Create several list for each correlated features */
List<List<FV>> correlatedFV = new ArrayList<>();
Set<FV> fvs = fv_list.keySet();
int[] selectedColumns = new int[CM.length];
for (int i = 0; i < CM.length; i++) {
for (int j = i + 1; j < CM.length; j++) {
/**
* For each correlated feature, put all FV into list and do selection based on top-k percent
*/
if (CM[i][j] >= corrThreshold) {
// Mark selectedColumns
selectedColumns[i] = 1;
selectedColumns[j] = 0;
// Add into list
List<FV> list = new ArrayList<>();
for (FV fv : fvs) {
if (fv.getFeature() == i || fv.getFeature() == j) {
list.add(fv);
}
}
selectSubsetTopKPercent(topk, correlatedFV, list);
}
}
}
/** Add remaining columns which are not selected (not correlated) */
for (int i = 0; i < selectedColumns.length; i++) {
if (selectedColumns[i] == 0) {
List<FV> list = new ArrayList<>();
for (FV fv : fvs) {
if (fv.getFeature() == i) {
list.add(fv);
}
}
selectSubsetTopKPercent(topk, correlatedFV, list);
}
}
/** Merge all correlated FV set and non-correlated FV set */
for (List<FV> list : correlatedFV) {
for (FV fv : list) {
result.put(fv, null);
}
}
/*
* TODO Correlation FVS 1. Create several list for each correlated
* features 2. For each correlated feature, put all FV into list and do
* selection based on top-k percent 3. When doing selection, check on
* every list (because every correlated feature would have different
* list and non-correlated feature, would have a separate list). For
* example: there is 5 features A, B, C, D, and E. Then, correlated
* pairs are: A with B, A with C, A with D, B with D, and C with D. Then
* we would have several list for them: AB list, AC list, AD list, BD
* list, CD list, and E list. (E is a non-correlated feature, thus have
* a separate list by itself). 4. Evaluate the parameters: top-k
* percents and threshold selection (for correlation).
*/
return result;
}
