// This takes an inordinate amount of time. -jdramsey 20150929
private int[] getNonMissingRows(Node x, Node y, List<Node> z) {
// List<Integer> rows = new ArrayList<Integer>();
//
// I:
// for (int i = 0; i < internalData.getNumRows(); i++) {
// for (Node node : variablesPerNode.get(x)) {
// if (isMissing(node, i)) continue I;
// }
//
// for (Node node : variablesPerNode.get(y)) {
// if (isMissing(node, i)) continue I;
// }
//
// for (Node _z : z) {
// for (Node node : variablesPerNode.get(_z)) {
// if (isMissing(node, i)) continue I;
// }
// }
//
// rows.add(i);
// }
// int[] _rows = new int[rows.size()];
// for (int k = 0; k < rows.size(); k++) _rows[k] = rows.get(k);
if (_rows == null) {
_rows = new int[internalData.getNumRows()];
for (int k = 0; k < _rows.length; k++) _rows[k] = k;
}
return _rows;
}
/**
* Constructs a test using a given data set. If a data set is provided (that is, a tabular data
* set), fourth moment statistics can be calculated (p. 160); otherwise, it must be assumed that
* the data are multivariate Gaussian.
*/
public DeltaSextadTest(DataSet dataSet) {
if (dataSet == null) {
throw new NullPointerException();
}
if (!dataSet.isContinuous()) {
throw new IllegalArgumentException();
}
this.cov = new CovarianceMatrix(dataSet);
List<DataSet> data1 = new ArrayList<DataSet>();
data1.add(dataSet);
List<DataSet> data2 = DataUtils.center(data1);
this.dataSet = data2.get(0);
this.data = this.dataSet.getDoubleData().transpose().toArray();
this.N = dataSet.getNumRows();
this.variables = dataSet.getVariables();
this.numVars = dataSet.getNumColumns();
this.variablesHash = new HashMap<Node, Integer>();
for (int i = 0; i < variables.size(); i++) {
variablesHash.put(variables.get(i), i);
}
this.means = new double[numVars];
for (int i = 0; i < numVars; i++) {
means[i] = mean(data[i], N);
}
}
/*
* @param dataSet A discrete data set.
* @param column the column in question.
* @return the max value in that column.
*/
private int maxInColumn(DataSet dataSet, int column) {
int max = -1;
for (int i = 0; i < dataSet.getNumRows(); i++) {
int value = dataSet.getInt(i, column);
if (value > max) max = value;
}
return max;
}
/** @return the splitNames selected by the editor. */
public static DataModel createSplits(DataSet dataSet, SplitCasesParams params) {
List<Integer> indices = new ArrayList<Integer>(dataSet.getNumRows());
for (int i = 0; i < dataSet.getNumRows(); i++) {
indices.add(i);
}
if (params.isDataShuffled()) {
Collections.shuffle(indices);
}
SplitCasesSpec spec = params.getSpec();
int numSplits = params.getNumSplits();
int sampleSize = spec.getSampleSize();
int[] breakpoints = spec.getBreakpoints();
List<String> splitNames = spec.getSplitNames();
int[] _breakpoints = new int[breakpoints.length + 2];
_breakpoints[0] = 0;
_breakpoints[_breakpoints.length - 1] = sampleSize;
System.arraycopy(breakpoints, 0, _breakpoints, 1, breakpoints.length);
DataModelList list = new DataModelList();
int ncols = dataSet.getNumColumns();
for (int n = 0; n < numSplits; n++) {
int _sampleSize = _breakpoints[n + 1] - _breakpoints[n];
DataSet _data = new ColtDataSet(_sampleSize, dataSet.getVariables());
_data.setName(splitNames.get(n));
for (int i = 0; i < _sampleSize; i++) {
int oldCase = indices.get(i + _breakpoints[n]);
for (int j = 0; j < ncols; j++) {
_data.setObject(i, j, dataSet.getObject(oldCase, j));
}
}
list.add(_data);
}
return list;
}
private void setDataSet(DataSet dataSet) {
List<String> _varNames = dataSet.getVariableNames();
this.variables = dataSet.getVariables();
this.dataSet = dataSet;
this.discrete = dataSet.isDiscrete();
if (!isDiscrete()) {
this.covariances = new CovarianceMatrix(dataSet);
}
this.sampleSize = dataSet.getNumRows();
}
/** Creates a cell count table for the given data set. */
public DataSetProbs(DataSet dataSet) {
if (dataSet == null) {
throw new NullPointerException();
}
this.dataSet = dataSet;
dims = new int[dataSet.getNumColumns()];
for (int i = 0; i < dims.length; i++) {
DiscreteVariable variable = (DiscreteVariable) dataSet.getVariable(i);
dims[i] = variable.getNumCategories();
}
numRows = dataSet.getNumRows();
}
private List<Node> expandVariable(DataSet dataSet, Node node) {
if (node instanceof ContinuousVariable) {
return Collections.singletonList(node);
}
if (node instanceof DiscreteVariable && ((DiscreteVariable) node).getNumCategories() < 3) {
return Collections.singletonList(node);
}
if (!(node instanceof DiscreteVariable)) {
throw new IllegalArgumentException();
}
List<String> varCats = new ArrayList<String>(((DiscreteVariable) node).getCategories());
// first category is reference
varCats.remove(0);
List<Node> variables = new ArrayList<Node>();
for (String cat : varCats) {
Node newVar;
do {
String newVarName = node.getName() + "MULTINOM" + "." + cat;
newVar = new DiscreteVariable(newVarName, 2);
} while (dataSet.getVariable(newVar.getName()) != null);
variables.add(newVar);
dataSet.addVariable(newVar);
int newVarIndex = dataSet.getColumn(newVar);
int numCases = dataSet.getNumRows();
for (int l = 0; l < numCases; l++) {
Object dataCell = dataSet.getObject(l, dataSet.getColumn(node));
int dataCellIndex = ((DiscreteVariable) node).getIndex(dataCell.toString());
if (dataCellIndex == ((DiscreteVariable) node).getIndex(cat))
dataSet.setInt(l, newVarIndex, 1);
else dataSet.setInt(l, newVarIndex, 0);
}
}
return variables;
}
private int sampleSize() {
return dataSet.getNumRows();
}
private double[] dependencePvalsLogit(Node x, Node y, List<Node> z) {
if (!variablesPerNode.containsKey(x)) {
throw new IllegalArgumentException("Unrecogized node: " + x);
}
if (!variablesPerNode.containsKey(y)) {
throw new IllegalArgumentException("Unrecogized node: " + y);
}
for (Node node : z) {
if (!variablesPerNode.containsKey(node)) {
throw new IllegalArgumentException("Unrecogized node: " + node);
}
}
List<Double> pValues = new ArrayList<Double>();
int[] _rows = getNonMissingRows(x, y, z);
logisticRegression.setRows(_rows);
List<Node> yzDumList = new ArrayList<>();
List<Node> yzList = new ArrayList<>();
yzList.add(y);
yzList.addAll(z);
// List<Node> zList = new ArrayList<>();
yzDumList.addAll(variablesPerNode.get(y));
for (Node _z : z) {
yzDumList.addAll(variablesPerNode.get(_z));
// zList.addAll(variablesPerNode.get(_z));
}
// double[][] coeffsDep = new double[variablesPerNode.get(x).size()][];
// DoubleMatrix2D coeffsNull = DoubleFactory2D.dense.make(zList.size(),
// variablesPerNode.get(x).size());
// DoubleMatrix2D coeffsDep = DoubleFactory2D.dense.make(yzDumList.size()+1,
// variablesPerNode.get(x).size());
double[] sumLnP = new double[yzList.size()];
for (int i = 0; i < sumLnP.length; i++) sumLnP[i] = 0.0;
for (int i = 0; i < variablesPerNode.get(x).size(); i++) {
Node _x = variablesPerNode.get(x).get(i);
LogisticRegression.Result result1 =
logisticRegression.regress((DiscreteVariable) _x, yzDumList);
int n = originalData.getNumRows();
int k = yzDumList.size();
// skip intercept at index 0
int coefIndex = 1;
for (int j = 0; j < yzList.size(); j++) {
for (int dum = 0; dum < variablesPerNode.get(yzList.get(j)).size(); dum++) {
double wald = Math.abs(result1.getCoefs()[coefIndex] / result1.getStdErrs()[coefIndex]);
// double val = (1.0 - new
// NormalDistribution(0,1).cumulativeProbability(wald))*2;//two-tailed test
// double val = 1-result1.getProbs()[i+1];
// this is exactly the same test as the linear case
double val = (1.0 - ProbUtils.tCdf(wald, n - k)) * 2;
// System.out.println(_x.getName() + "\t" + yzDumList.get(coefIndex-1).getName() + "\t" +
// val + "\t" + (n-k));
// if(val <= 0) System.out.println("Zero p-val t-test: p " + val + " stat " + wald + " k "
// + k + " n " + n);
sumLnP[j] += Math.log(val);
coefIndex++;
}
}
}
double[] pVec = new double[sumLnP.length];
for (int i = 0; i < pVec.length; i++) {
if (sumLnP[i] == Double.NEGATIVE_INFINITY) pVec[i] = 0.0;
else {
int df = 2 * variablesPerNode.get(x).size() * variablesPerNode.get(yzList.get(i)).size();
pVec[i] = 1.0 - new ChiSquaredDistribution(df).cumulativeProbability(-2 * sumLnP[i]);
}
}
return pVec;
}
private void createDiscreteTimeSeriesData() {
// GIVEN: Continuous data set D, maximum lag m.
Node[] dataVars = dataSet.getVariables().toArray(new Node[0]);
int n = dataVars.length;
int m = getNumLags();
// LetXi, i = 0,...,n-1, be the variables from the data. Let Xi(t) be
// the variable Xi at time lag t (before 0), t = 0,...,m.
Node[][] laggedVars = new Node[m + 1][n];
Knowledge knowledge = new Knowledge();
for (int s = 0; s <= m; s++) {
for (int j = 0; j < n; j++) {
String name1 = dataVars[j].getName();
String name2 = name1 + "." + (s + 1);
laggedVars[s][j] = new DiscreteVariable((DiscreteVariable) dataVars[j]);
laggedVars[s][j].setName(name2);
laggedVars[s][j].setCenter(80 * j + 50, 80 * (m - s) + 50);
knowledge.addToTier(s, laggedVars[s][j].getName());
}
}
// 2. Prepare the data the way you did.
List<Node> variables = new LinkedList<Node>();
for (int s = 0; s <= m; s++) {
for (int i = 0; i < n; i++) {
int[] rawData = new int[dataSet.getNumRows()];
for (int j = 0; j < dataSet.getNumRows(); j++) {
rawData[j] = dataSet.getInt(j, i);
}
int size = dataSet.getNumRows();
int[] laggedRaw = new int[size - m + 1];
System.arraycopy(rawData, m - s, laggedRaw, 0, size - m + 1);
variables.add(laggedVars[s][i]);
}
}
DataSet _laggedData = new ColtDataSet(dataSet.getNumRows() - m + 1, variables);
for (int s = 0; s <= m; s++) {
for (int i = 0; i < n; i++) {
int[] rawData = new int[dataSet.getNumRows()];
for (int j = 0; j < dataSet.getNumRows(); j++) {
rawData[j] = dataSet.getInt(j, i);
}
int size = dataSet.getNumRows();
int[] laggedRaw = new int[size - m + 1];
System.arraycopy(rawData, m - s, laggedRaw, 0, size - m + 1);
int _col = _laggedData.getColumn(laggedVars[s][i]);
for (int j = 0; j < dataSet.getNumRows(); j++) {
_laggedData.setInt(j, _col, laggedRaw[j]);
}
}
}
knowledge.setDefaultToKnowledgeLayout(true);
_laggedData.setKnowledge(knowledge);
DataModelList list = new DataModelList();
list.add(_laggedData);
getDataEditor().reset(list);
getDataEditor().selectLastTab();
}
protected SemIm estimateCoeffs(SemIm semIm) {
// System.out.print("\n****************\nCalling 2SLS... ");
SemGraph semGraph = semIm.getSemPm().getGraph();
// Get list of fixed measurements that will be kept fixed, and the
// respective latent variables that are their parents.
// "X" variables are exogenous, while "Y" variables are endogenous.
List<Node> ly = new LinkedList<Node>();
List<Node> lx = new LinkedList<Node>();
List<Node> my1 = new LinkedList<Node>();
List<Node> mx1 = new LinkedList<Node>();
List<Node> observed = new LinkedList<Node>();
for (Node nodeA : semGraph.getNodes()) {
if (nodeA.getNodeType() == NodeType.ERROR) {
continue;
}
if (nodeA.getNodeType() == NodeType.LATENT) {
if (semGraph.getParents(nodeA).size() == 0) {
lx.add(nodeA);
} else {
ly.add(nodeA);
}
} else {
observed.add(nodeA);
}
}
setFixedNodes(semGraph, mx1, my1);
// ------------------------------------------------------------------
// Estimate freeParameters for the latent/latent edges
for (Node current : ly) {
if (nodeName != null && !nodeName.equals(current.getName())) {
continue;
}
// Build Z, the matrix containing the data for the fixed measurements
// associated with the parents of the getModel (endogenous) latent node
List<Node> endo_parents_m = new LinkedList<Node>();
List<Node> exo_parents_m = new LinkedList<Node>();
List<Node> endo_parents = new LinkedList<Node>();
List<Node> exo_parents = new LinkedList<Node>();
Iterator<Node> it_p = semGraph.getParents(current).iterator();
lNames = new String[lx.size() + ly.size()];
while (it_p.hasNext()) {
Node node = it_p.next();
if (node.getNodeType() == NodeType.ERROR) {
continue;
}
if (lx.contains(node)) {
int position = lx.indexOf(node);
exo_parents_m.add(mx1.get(position));
exo_parents.add(node);
} else {
int position = ly.indexOf(node);
endo_parents_m.add(my1.get(position));
endo_parents.add(node);
}
}
Object endp_a_m[] = endo_parents_m.toArray();
Object exop_a_m[] = exo_parents_m.toArray();
Object endp_a[] = endo_parents.toArray();
Object exop_a[] = exo_parents.toArray();
int n = dataSet.getNumRows(), c = endp_a_m.length + exop_a_m.length;
if (c == 0) {
continue;
}
double Z[][] = new double[n][c];
int count = 0;
for (int i = 0; i < endp_a_m.length; i++) {
Node node = (Node) endp_a_m[i];
String name = node.getName();
Node variable = dataSet.getVariable(name);
int colIndex = dataSet.getVariables().indexOf(variable);
// Column column = dataSet.getColumnObject(variable);
// double column_data[] = (double[]) column.getRawData();
for (int j = 0; j < n; j++) {
// Z[j][i] = column_data[j];
Z[j][i] = dataSet.getDouble(j, colIndex);
}
lNames[count++] = (endo_parents.get(i)).getName();
}
for (int i = 0; i < exop_a_m.length; i++) {
Node node = (Node) exop_a_m[i];
String name = node.getName();
Node variable = dataSet.getVariable(name);
int colIndex = dataSet.getVariables().indexOf(variable);
// Column column = dataSet.getColumnObject(variable);
// double column_data[] = (double[]) column.getRawData();
for (int j = 0; j < n; j++) {
// Z[j][endp_a_m.length + i] = column_data[j];
Z[j][endp_a_m.length + i] = dataSet.getDouble(j, colIndex);
}
lNames[count++] = exo_parents.get(i).getName();
}
// Build V, the matrix containing the data for the nonfixed measurements
// associated with the parents of the getModel (endogenous) latent node
endo_parents_m = new LinkedList<Node>();
exo_parents_m = new LinkedList<Node>();
it_p = semGraph.getParents(current).iterator();
while (it_p.hasNext()) {
Node node = it_p.next();
if (node.getNodeType() == NodeType.ERROR) {
continue;
}
List<Node> other_measures = new LinkedList<Node>();
for (Node next : semGraph.getChildren(node)) {
if (next.getNodeType() == NodeType.MEASURED) {
other_measures.add(next);
}
}
if (lx.contains(node)) {
int position = lx.indexOf(node);
other_measures.remove(mx1.get(position));
exo_parents_m.addAll(other_measures);
} else {
int position = ly.indexOf(node);
other_measures.remove(my1.get(position));
endo_parents_m.addAll(other_measures);
}
}
endp_a_m = endo_parents_m.toArray();
exop_a_m = exo_parents_m.toArray();
n = dataSet.getNumRows();
c = endp_a_m.length + exop_a_m.length;
double V[][] = new double[n][c];
if (c == 0) {
continue;
}
for (int i = 0; i < endp_a_m.length; i++) {
Node node = ((Node) endp_a_m[i]);
String name = node.getName();
Node variable = dataSet.getVariable(name);
int colIndex = dataSet.getVariables().indexOf(variable);
// Column column = dataSet.getColumnObject(variable);
// double column_data[] = (double[]) column.getRawData();
for (int j = 0; j < n; j++) {
// V[j][i] = column_data[j];
V[j][i] = dataSet.getDouble(j, colIndex);
}
}
for (int i = 0; i < exop_a_m.length; i++) {
Node node = (Node) exop_a_m[i];
String name = node.getName();
Node variable = dataSet.getVariable(name);
int colIndex = dataSet.getVariables().indexOf(variable);
// Column column = dataSet.getColumnObject(variable);
// double column_data[] = (double[]) column.getRawData();
for (int j = 0; j < n; j++) {
// V[j][endp_a_m.length + i] = column_data[j];
V[j][endp_a_m.length + i] = dataSet.getDouble(j, colIndex);
}
}
double yi[] = new double[n];
if (lx.contains(current)) {
int position = lx.indexOf(current);
Node node = mx1.get(position);
String name = node.getName();
Node variable = dataSet.getVariable(name);
int colIndex = dataSet.getVariables().indexOf(variable);
// Column column = dataSet.getColumnObject(variable);
//
// System.arraycopy(column.getRawData(), 0, yi, 0, n);
for (int i = 0; i < n; i++) {
yi[i] = dataSet.getDouble(i, colIndex);
}
} else {
int position = ly.indexOf(current);
Node node = my1.get(position);
String name = node.getName();
Node variable = dataSet.getVariable(name);
int colIndex = dataSet.getVariables().indexOf(variable);
// System.arraycopy(dataSet.getColumnObject(variable).getRawData(), 0, yi, 0,
// n);
for (int i = 0; i < n; i++) {
yi[i] = dataSet.getDouble(i, colIndex);
}
}
// Build Z_hat
double Z_hat[][] =
MatrixUtils.product(
V,
MatrixUtils.product(
MatrixUtils.inverse(MatrixUtils.product(MatrixUtils.transpose(V), V)),
MatrixUtils.product(MatrixUtils.transpose(V), Z)));
A_hat =
MatrixUtils.product(
MatrixUtils.inverse(MatrixUtils.product(MatrixUtils.transpose(Z_hat), Z_hat)),
MatrixUtils.product(MatrixUtils.transpose(Z_hat), yi));
// Set the edge for the fixed measurement
int position = ly.indexOf(current);
semIm.setParamValue(current, my1.get(position), 1.);
// Set the edge for the latents
for (int i = 0; i < endp_a.length; i++) {
semIm.setParamValue((Node) endp_a[i], current, A_hat[i]);
}
for (int i = 0; i < exop_a.length; i++) {
semIm.setParamValue((Node) exop_a[i], current, A_hat[endp_a.length + i]);
}
if (nodeName != null && nodeName.equals(current.getName())) {
computeAsymptLatentCovar(yi, A_hat, Z, Z_hat, dataSet.getNumRows());
break;
}
}
// ------------------------------------------------------------------
// Estimate freeParameters of the measurement model
// Set the edges of the fixed measurements of exogenous
for (Node current : lx) {
int position = lx.indexOf(current);
semIm.setParamValue(current, mx1.get(position), 1.);
}
for (Node current : observed) {
if (nodeName != null && !nodeName.equals(current.getName())) {
continue;
}
if (mx1.contains(current) || my1.contains(current)) {
continue;
}
// First, get the parent of this observed
Node current_latent = null;
for (Node node : semGraph.getParents(current)) {
if (node.getNodeType() == NodeType.ERROR) {
continue;
}
current_latent = node;
}
Iterator<Node> children = semGraph.getChildren(current_latent).iterator();
List<Node> other_measures = new LinkedList<Node>();
Node fixed_measurement;
while (children.hasNext()) {
Node next = children.next();
if ((next.getNodeType() == NodeType.MEASURED) && next != current) {
other_measures.add(next);
}
}
if (lx.contains(current_latent)) {
int position = lx.indexOf(current_latent);
other_measures.remove(mx1.get(position));
fixed_measurement = mx1.get(position);
} else {
int position = ly.indexOf(current_latent);
other_measures.remove(my1.get(position));
fixed_measurement = my1.get(position);
}
// Regress other_measures over the fixed measurement x1 (y1) correspondent
// to the measurement variable that is being evaluated
int n = dataSet.getNumRows(), c = other_measures.size();
if (c == 0) {
continue;
}
double Z[][] = new double[n][c];
for (int i = 0; i < c; i++) {
Node variable = dataSet.getVariable((other_measures.get(i)).getName());
int varIndex = dataSet.getVariables().indexOf(variable);
// Column column = dataSet.getColumnObject(variable);
// double column_data[] = (double[]) column.getRawData();
for (int j = 0; j < n; j++) {
// Z[j][i] = column_data[j];
Z[j][i] = dataSet.getDouble(varIndex, j);
}
}
// Build C, the column matrix containing the data for the fixed
// measurement associated with the only latent parent of the getModel
// observed node (as assumed by the structure of our measurement model).
Node variable = dataSet.getVariable(fixed_measurement.getName());
int colIndex = dataSet.getVariables().indexOf(variable);
// Column column = dataSet.getColumnObject(variable);
// double C[] = (double[]) column.getRawData();
double[] C = new double[dataSet.getNumRows()];
for (int i = 0; i < dataSet.getNumRows(); i++) {
C[i] = dataSet.getDouble(colIndex, i);
}
// Build V, the matrix containing the data for the other measurements
// associated with the parents of the (latent) parent of getModel
// observed node. The only difference with respect to the estimation
// of the within-latent coefficients is that here we only include
// the other measurements attached to the parent of the getModel node,
// assuming that the error term of the getModel node is independent
// of the error term of the others and that each measurement is
// taken with respect to only one latent.
n = dataSet.getNumRows();
c = other_measures.size();
double V[][] = new double[n][c];
for (int i = 0; i < c; i++) {
Node variable2 = dataSet.getVariable((other_measures.get(i)).getName());
int var2index = dataSet.getVariables().indexOf(variable2);
// Column column = dataSet.getColumnObject(variable2);
// double column_data[] = (double[]) column.getRawData();
for (int j = 0; j < n; j++) {
// V[j][i] = column_data[j];
V[j][i] = dataSet.getDouble(j, var2index);
}
}
double yi[] = new double[n];
Node variable3 = dataSet.getVariable((current).getName());
int var3Index = dataSet.getVariables().indexOf(variable3);
for (int i = 0; i < n; i++) {
yi[i] = dataSet.getDouble(i, var3Index);
}
// Object rawData = dataSet.getColumnObject(variable3).getRawData();
// System.arraycopy(rawData, 0, yi, 0, n);
double C_hat[] =
MatrixUtils.product(
V,
MatrixUtils.product(
MatrixUtils.inverse(MatrixUtils.product(MatrixUtils.transpose(V), V)),
MatrixUtils.product(MatrixUtils.transpose(V), C)));
double A_hat =
MatrixUtils.innerProduct(
MatrixUtils.scalarProduct(1. / MatrixUtils.innerProduct(C_hat, C_hat), C_hat), yi);
// Set the edge for the getModel measurement
semIm.setParamValue(current_latent, current, A_hat);
}
return semIm;
}
/**
* This method takes an instantiated Bayes net (BayesIm) whose graph include all the variables
* (observed and latent) and computes estimated counts using the data in the DataSet mixedData.
* The counts that are estimated correspond to cells in the conditional probability tables of the
* Bayes net. The outermost loop (indexed by j) is over the set of variables. If the variable has
* no parents, each case in the dataset is examined and the count for the observed value of the
* variables is increased by 1.0; if the value of the variable is missing the marginal
* probabilities its values given the values of the variables that are available for that case are
* used to increment the corresponding estimated counts. If a variable has parents then there is a
* loop which steps through all possible sets of values of its parents. This loop is indexed by
* the variable "row". Each case in the dataset is examined. It the variable and all its parents
* have values in the case the corresponding estimated counts are incremented by 1.0. If the
* variable or any of its parents have missing values, the joint marginal is computed for the
* variable and the set of values of its parents corresponding to "row" and the corresponding
* estimated counts are incremented by the appropriate probability. The estimated counts are
* stored in the double[][][] array estimatedCounts. The count (possibly fractional) of the number
* of times each combination of parent values occurs is stored in the double[][] array
* estimatedCountsDenom. These two arrays are used to compute the estimated conditional
* probabilities of the output Bayes net.
*/
private BayesIm expectation(BayesIm inputBayesIm) {
// System.out.println("Entered method expectation.");
int numCases = mixedData.getNumRows();
// StoredCellEstCounts estCounts = new StoredCellEstCounts(variables);
int numVariables = allVariables.size();
RowSummingExactUpdater rseu = new RowSummingExactUpdater(inputBayesIm);
for (int j = 0; j < numVariables; j++) {
DiscreteVariable var = (DiscreteVariable) allVariables.get(j);
String varName = var.getName();
Node varNode = graph.getNode(varName);
int varIndex = inputBayesIm.getNodeIndex(varNode);
int[] parentVarIndices = inputBayesIm.getParents(varIndex);
// System.out.println("graph = " + graph);
// for(int col = 0; col < var.getNumSplits(); col++)
// System.out.println("Category " + col + " = " + var.getCategory(col));
// System.out.println("Updating estimated counts for node " + varName);
// This segment is for variables with no parents:
if (parentVarIndices.length == 0) {
// System.out.println("No parents");
for (int col = 0; col < var.getNumCategories(); col++) {
estimatedCounts[j][0][col] = 0.0;
}
for (int i = 0; i < numCases; i++) {
// System.out.println("Case " + i);
// If this case has a value for var
if (mixedData.getInt(i, j) != -99) {
estimatedCounts[j][0][mixedData.getInt(i, j)] += 1.0;
// System.out.println("Adding 1.0 to " + varName +
// " row 0 category " + mixedData[j][i]);
} else {
// find marginal probability, given obs data in this case, p(v=0)
Evidence evidenceThisCase = Evidence.tautology(inputBayesIm);
boolean existsEvidence = false;
// Define evidence for updating by using the values of the other vars.
for (int k = 0; k < numVariables; k++) {
if (k == j) {
continue;
}
Node otherVar = allVariables.get(k);
if (mixedData.getInt(i, k) == -99) {
continue;
}
existsEvidence = true;
String otherVarName = otherVar.getName();
Node otherNode = graph.getNode(otherVarName);
int otherIndex = inputBayesIm.getNodeIndex(otherNode);
evidenceThisCase.getProposition().setCategory(otherIndex, mixedData.getInt(i, k));
}
if (!existsEvidence) {
continue; // No other variable contained useful data
}
rseu.setEvidence(evidenceThisCase);
for (int m = 0; m < var.getNumCategories(); m++) {
estimatedCounts[j][0][m] += rseu.getMarginal(varIndex, m);
// System.out.println("Adding " + p + " to " + varName +
// " row 0 category " + m);
// find marginal probability, given obs data in this case, p(v=1)
// estimatedCounts[j][0][1] += 0.5;
}
}
}
// Print estimated counts:
// System.out.println("Estimated counts: ");
// Print counts for each value of this variable with no parents.
// for(int m = 0; m < var.getNumSplits(); m++)
// System.out.print(" " + m + " " + estimatedCounts[j][0][m]);
// System.out.println();
} else { // For variables with parents:
int numRows = inputBayesIm.getNumRows(varIndex);
for (int row = 0; row < numRows; row++) {
int[] parValues = inputBayesIm.getParentValues(varIndex, row);
estimatedCountsDenom[varIndex][row] = 0.0;
for (int col = 0; col < var.getNumCategories(); col++) {
estimatedCounts[varIndex][row][col] = 0.0;
}
for (int i = 0; i < numCases; i++) {
// for a case where the parent values = parValues increment the estCount
boolean parentMatch = true;
for (int p = 0; p < parentVarIndices.length; p++) {
if (parValues[p] != mixedData.getInt(i, parentVarIndices[p])
&& mixedData.getInt(i, parentVarIndices[p]) != -99) {
parentMatch = false;
break;
}
}
if (!parentMatch) {
continue; // Not a matching case; go to next.
}
boolean parentMissing = false;
for (int parentVarIndice : parentVarIndices) {
if (mixedData.getInt(i, parentVarIndice) == -99) {
parentMissing = true;
break;
}
}
if (mixedData.getInt(i, j) != -99 && !parentMissing) {
estimatedCounts[j][row][mixedData.getInt(i, j)] += 1.0;
estimatedCountsDenom[j][row] += 1.0;
continue; // Next case
}
// for a case with missing data (either var or one of its parents)
// compute the joint marginal
// distribution for var & this combination of values of its parents
// and update the estCounts accordingly
// To compute marginals create the evidence
boolean existsEvidence = false;
Evidence evidenceThisCase = Evidence.tautology(inputBayesIm);
// "evidenceVars" not used.
// List<String> evidenceVars = new LinkedList<String>();
// for (int k = 0; k < numVariables; k++) {
// //if(k == j) continue;
// Variable otherVar = allVariables.get(k);
// if (mixedData.getInt(i, k) == -99) {
// continue;
// }
// existsEvidence = true;
// String otherVarName = otherVar.getName();
// Node otherNode = graph.getNode(otherVarName);
// int otherIndex = inputBayesIm.getNodeIndex(
// otherNode);
// evidenceThisCase.getProposition().setCategory(
// otherIndex, mixedData.getInt(i, k));
// evidenceVars.add(otherVarName);
// }
if (!existsEvidence) {
continue;
}
rseu.setEvidence(evidenceThisCase);
estimatedCountsDenom[j][row] += rseu.getJointMarginal(parentVarIndices, parValues);
int[] parPlusChildIndices = new int[parentVarIndices.length + 1];
int[] parPlusChildValues = new int[parentVarIndices.length + 1];
parPlusChildIndices[0] = varIndex;
for (int pc = 1; pc < parPlusChildIndices.length; pc++) {
parPlusChildIndices[pc] = parentVarIndices[pc - 1];
parPlusChildValues[pc] = parValues[pc - 1];
}
for (int m = 0; m < var.getNumCategories(); m++) {
parPlusChildValues[0] = m;
/*
if(varName.equals("X1") && i == 0 ) {
System.out.println("Calling getJointMarginal with parvalues");
for(int k = 0; k < parPlusChildIndices.length; k++) {
int pIndex = parPlusChildIndices[k];
Node pNode = inputBayesIm.getNode(pIndex);
String pName = pNode.getName();
System.out.println(pName + " " + parPlusChildValues[k]);
}
}
*/
/*
if(varName.equals("X1") && i == 0 ) {
System.out.println("Evidence = " + evidenceThisCase);
//int[] vars = {l1Index, x1Index};
Node nodex1 = inputBayesIm.getNode("X1");
int x1Index = inputBayesIm.getNodeIndex(nodex1);
Node nodel1 = inputBayesIm.getNode("L1");
int l1Index = inputBayesIm.getNodeIndex(nodel1);
int[] vars = {l1Index, x1Index};
int[] vals = {0, 0};
double ptest = rseu.getJointMarginal(vars, vals);
System.out.println("Joint marginal (X1=0, L1 = 0) = " + p);
}
*/
estimatedCounts[j][row][m] +=
rseu.getJointMarginal(parPlusChildIndices, parPlusChildValues);
// System.out.println("Case " + i + " parent values ");
// for (int pp = 0; pp < parentVarIndices.length; pp++) {
// Variable par = (Variable) allVariables.get(parentVarIndices[pp]);
// System.out.print(" " + par.getName() + " " + parValues[pp]);
// }
// System.out.println();
// System.out.println("Adding " + p + " to " + varName +
// " row " + row + " category " + m);
}
// }
}
// Print estimated counts:
// System.out.println("Estimated counts: ");
// System.out.println(" Parent values: ");
// for (int i = 0; i < parentVarIndices.length; i++) {
// Variable par = (Variable) allVariables.get(parentVarIndices[i]);
// System.out.print(" " + par.getName() + " " + parValues[i] + " ");
// }
// System.out.println();
// for(int m = 0; m < var.getNumSplits(); m++)
// System.out.print(" " + m + " " + estimatedCounts[j][row][m]);
// System.out.println();
}
} // else
} // j < numVariables
BayesIm outputBayesIm = new MlBayesIm(bayesPm);
for (int j = 0; j < nodes.length; j++) {
DiscreteVariable var = (DiscreteVariable) allVariables.get(j);
String varName = var.getName();
Node varNode = graph.getNode(varName);
int varIndex = inputBayesIm.getNodeIndex(varNode);
// int[] parentVarIndices = inputBayesIm.getParents(varIndex);
int numRows = inputBayesIm.getNumRows(j);
// System.out.println("Conditional probabilities for variable " + varName);
int numCols = inputBayesIm.getNumColumns(j);
if (numRows == 1) {
double sum = 0.0;
for (int m = 0; m < numCols; m++) {
sum += estimatedCounts[j][0][m];
}
for (int m = 0; m < numCols; m++) {
condProbs[j][0][m] = estimatedCounts[j][0][m] / sum;
// System.out.print(" " + condProbs[j][0][m]);
outputBayesIm.setProbability(varIndex, 0, m, condProbs[j][0][m]);
}
// System.out.println();
} else {
for (int row = 0; row < numRows; row++) {
// int[] parValues = inputBayesIm.getParentValues(varIndex,
// row);
// int numCols = inputBayesIm.getNumColumns(j);
// for (int p = 0; p < parentVarIndices.length; p++) {
// Variable par = (Variable) allVariables.get(parentVarIndices[p]);
// System.out.print(" " + par.getName() + " " + parValues[p]);
// }
// double sum = 0.0;
// for(int m = 0; m < numCols; m++)
// sum += estimatedCounts[j][row][m];
for (int m = 0; m < numCols; m++) {
if (estimatedCountsDenom[j][row] != 0.0) {
condProbs[j][row][m] = estimatedCounts[j][row][m] / estimatedCountsDenom[j][row];
} else {
condProbs[j][row][m] = Double.NaN;
}
// System.out.print(" " + condProbs[j][row][m]);
outputBayesIm.setProbability(varIndex, row, m, condProbs[j][row][m]);
}
// System.out.println();
}
}
}
return outputBayesIm;
}
private void initialize() {
DirichletBayesIm prior = DirichletBayesIm.symmetricDirichletIm(bayesPmObs, 0.5);
observedIm = DirichletEstimator.estimate(prior, dataSet);
// MLBayesEstimator dirichEst = new MLBayesEstimator();
// observedIm = dirichEst.estimate(bayesPmObs, dataSet);
// System.out.println("Estimated Bayes IM for Measured Variables: ");
// System.out.println(observedIm);
// mixedData should be ddsNm with new columns for the latent variables.
// Each such column should contain missing data for each case.
int numFullCases = dataSet.getNumRows();
List<Node> variables = new LinkedList<Node>();
for (Node node : nodes) {
if (node.getNodeType() == NodeType.LATENT) {
int numCategories = bayesPm.getNumCategories(node);
DiscreteVariable latentVar = new DiscreteVariable(node.getName(), numCategories);
variables.add(latentVar);
} else {
String name = bayesPm.getVariable(node).getName();
Node variable = dataSet.getVariable(name);
variables.add(variable);
}
}
DataSet dsMixed = new ColtDataSet(numFullCases, variables);
for (int j = 0; j < nodes.length; j++) {
if (nodes[j].getNodeType() == NodeType.LATENT) {
for (int i = 0; i < numFullCases; i++) {
dsMixed.setInt(i, j, -99);
}
} else {
String name = bayesPm.getVariable(nodes[j]).getName();
Node variable = dataSet.getVariable(name);
int index = dataSet.getColumn(variable);
for (int i = 0; i < numFullCases; i++) {
dsMixed.setInt(i, j, dataSet.getInt(i, index));
}
}
}
// System.out.println(dsMixed);
mixedData = dsMixed;
allVariables = mixedData.getVariables();
// Find the bayes net which is parameterized using mixedData or set randomly when that's
// not possible.
estimateIM(bayesPm, mixedData);
// The following DEBUG section tests a case specified by P. Spirtes
// DEBUG TAIL: For use with embayes_l1x1x2x3V3.dat
/*
Node l1Node = graph.getNode("L1");
//int l1Index = bayesImMixed.getNodeIndex(l1Node);
int l1index = estimatedIm.getNodeIndex(l1Node);
Node x1Node = graph.getNode("X1");
//int x1Index = bayesImMixed.getNodeIndex(x1Node);
int x1Index = estimatedIm.getNodeIndex(x1Node);
Node x2Node = graph.getNode("X2");
//int x2Index = bayesImMixed.getNodeIndex(x2Node);
int x2Index = estimatedIm.getNodeIndex(x2Node);
Node x3Node = graph.getNode("X3");
//int x3Index = bayesImMixed.getNodeIndex(x3Node);
int x3Index = estimatedIm.getNodeIndex(x3Node);
estimatedIm.setProbability(l1index, 0, 0, 0.5);
estimatedIm.setProbability(l1index, 0, 1, 0.5);
//bayesImMixed.setProbability(x1Index, 0, 0, 0.33333);
//bayesImMixed.setProbability(x1Index, 0, 1, 0.66667);
estimatedIm.setProbability(x1Index, 0, 0, 0.6); //p(x1 = 0 | l1 = 0)
estimatedIm.setProbability(x1Index, 0, 1, 0.4); //p(x1 = 1 | l1 = 0)
estimatedIm.setProbability(x1Index, 1, 0, 0.4); //p(x1 = 0 | l1 = 1)
estimatedIm.setProbability(x1Index, 1, 1, 0.6); //p(x1 = 1 | l1 = 1)
//bayesImMixed.setProbability(x2Index, 1, 0, 0.66667);
//bayesImMixed.setProbability(x2Index, 1, 1, 0.33333);
estimatedIm.setProbability(x2Index, 1, 0, 0.4); //p(x2 = 0 | l1 = 1)
estimatedIm.setProbability(x2Index, 1, 1, 0.6); //p(x2 = 1 | l1 = 1)
estimatedIm.setProbability(x2Index, 0, 0, 0.6); //p(x2 = 0 | l1 = 0)
estimatedIm.setProbability(x2Index, 0, 1, 0.4); //p(x2 = 1 | l1 = 0)
//bayesImMixed.setProbability(x3Index, 1, 0, 0.66667);
//bayesImMixed.setProbability(x3Index, 1, 1, 0.33333);
estimatedIm.setProbability(x3Index, 1, 0, 0.4); //p(x3 = 0 | l1 = 1)
estimatedIm.setProbability(x3Index, 1, 1, 0.6); //p(x3 = 1 | l1 = 1)
estimatedIm.setProbability(x3Index, 0, 0, 0.6); //p(x3 = 0 | l1 = 0)
estimatedIm.setProbability(x3Index, 0, 1, 0.4); //p(x3 = 1 | l1 = 0)
*/
// END of TAIL
// System.out.println("bayes IM estimated by estimateIM");
// System.out.println(bayesImMixed);
// System.out.println(estimatedIm);
estimatedCounts = new double[nodes.length][][];
estimatedCountsDenom = new double[nodes.length][];
condProbs = new double[nodes.length][][];
for (int i = 0; i < nodes.length; i++) {
// int numRows = bayesImMixed.getNumRows(i);
int numRows = estimatedIm.getNumRows(i);
estimatedCounts[i] = new double[numRows][];
estimatedCountsDenom[i] = new double[numRows];
condProbs[i] = new double[numRows][];
// for(int j = 0; j < bayesImMixed.getNumRows(i); j++) {
for (int j = 0; j < estimatedIm.getNumRows(i); j++) {
// int numCols = bayesImMixed.getNumColumns(i);
int numCols = estimatedIm.getNumColumns(i);
estimatedCounts[i][j] = new double[numCols];
condProbs[i][j] = new double[numCols];
}
}
}
public final DataSet filter(DataSet dataSet) {
// Why does it have to be discrete? Why can't we simply expand
// whatever discrete columns are there and leave the continuous
// ones untouched? jdramsey 7/4/2005
// if (!(dataSet.isDiscrete())) {
// throw new IllegalArgumentException("Data set must be discrete.");
// }
List<Node> variables = new LinkedList<>();
// Add all of the variables to the new data set.
for (int j = 0; j < dataSet.getNumColumns(); j++) {
Node _var = dataSet.getVariable(j);
if (!(_var instanceof DiscreteVariable)) {
variables.add(_var);
continue;
}
DiscreteVariable variable = (DiscreteVariable) _var;
String oldName = variable.getName();
List<String> oldCategories = variable.getCategories();
List<String> newCategories = new LinkedList<>(oldCategories);
String newCategory = "Missing";
int _j = 0;
while (oldCategories.contains(newCategory)) {
newCategory = "Missing" + (++_j);
}
newCategories.add(newCategory);
String newName = oldName + "+";
DiscreteVariable newVariable = new DiscreteVariable(newName, newCategories);
variables.add(newVariable);
}
DataSet newDataSet = new ColtDataSet(dataSet.getNumRows(), variables);
// Copy old values to new data set, replacing missing values with new
// "MissingValue" categories.
for (int j = 0; j < dataSet.getNumColumns(); j++) {
Node _var = dataSet.getVariable(j);
if (_var instanceof ContinuousVariable) {
for (int i = 0; i < dataSet.getNumRows(); i++) {
newDataSet.setDouble(i, j, dataSet.getDouble(i, j));
}
} else if (_var instanceof DiscreteVariable) {
DiscreteVariable variable = (DiscreteVariable) _var;
int numCategories = variable.getNumCategories();
for (int i = 0; i < dataSet.getNumRows(); i++) {
int value = dataSet.getInt(i, j);
if (value == DiscreteVariable.MISSING_VALUE) {
newDataSet.setInt(i, j, numCategories);
} else {
newDataSet.setInt(i, j, value);
}
}
}
}
return newDataSet;
}