private void calculateArrowsForward(Node x, Node y, Graph graph) {
clearArrow(x, y);
if (!knowledgeEmpty()) {
if (getKnowledge().isForbidden(x.getName(), y.getName())) {
return;
}
}
List<Node> naYX = getNaYX(x, y, graph);
List<Node> t = getTNeighbors(x, y, graph);
DepthChoiceGenerator gen = new DepthChoiceGenerator(t.size(), t.size());
int[] choice;
while ((choice = gen.next()) != null) {
List<Node> s = GraphUtils.asList(choice, t);
if (!knowledgeEmpty()) {
if (!validSetByKnowledge(y, s)) {
continue;
}
}
double bump = insertEval(x, y, s, naYX, graph);
if (bump > 0.0) {
Arrow arrow = new Arrow(bump, x, y, s, naYX);
sortedArrows.add(arrow);
addLookupArrow(x, y, arrow);
}
}
}
/**
* Use background knowledge to decide if an insert or delete operation does not orient edges in a
* forbidden direction according to prior knowledge. If some orientation is forbidden in the
* subset, the whole subset is forbidden.
*/
private boolean validSetByKnowledge(Node y, List<Node> subset) {
for (Node node : subset) {
if (getKnowledge().isForbidden(node.getName(), y.getName())) {
return false;
}
}
return true;
}
private Node getVariable(List<Node> variables, String name) {
for (Node node : variables) {
if (name.equals(node.getName())) {
return node;
}
}
return null;
}
private double pValue(Node node, List<Node> parents) {
List<Double> _residuals = new ArrayList<Double>();
Node _target = node;
List<Node> _regressors = parents;
Node target = getVariable(variables, _target.getName());
List<Node> regressors = new ArrayList<Node>();
for (Node _regressor : _regressors) {
Node variable = getVariable(variables, _regressor.getName());
regressors.add(variable);
}
DATASET:
for (int m = 0; m < dataSets.size(); m++) {
RegressionResult result = regressions.get(m).regress(target, regressors);
TetradVector residualsSingleDataset = result.getResiduals();
for (int h = 0; h < residualsSingleDataset.size(); h++) {
if (Double.isNaN(residualsSingleDataset.get(h))) {
continue DATASET;
}
}
DoubleArrayList _residualsSingleDataset =
new DoubleArrayList(residualsSingleDataset.toArray());
double mean = Descriptive.mean(_residualsSingleDataset);
double std =
Descriptive.standardDeviation(
Descriptive.variance(
_residualsSingleDataset.size(),
Descriptive.sum(_residualsSingleDataset),
Descriptive.sumOfSquares(_residualsSingleDataset)));
for (int i2 = 0; i2 < _residualsSingleDataset.size(); i2++) {
// _residualsSingleDataset.set(i2, (_residualsSingleDataset.get(i2) - mean) /
// std);
if (isMeanCenterResiduals()) {
_residualsSingleDataset.set(i2, (_residualsSingleDataset.get(i2) - mean));
}
// _residualsSingleDataset.set(i2, (_residualsSingleDataset.get(i2)));
}
for (int k = 0; k < _residualsSingleDataset.size(); k++) {
_residuals.add(_residualsSingleDataset.get(k));
}
}
double[] _f = new double[_residuals.size()];
for (int k = 0; k < _residuals.size(); k++) {
_f[k] = _residuals.get(k);
}
return new AndersonDarlingTest(_f).getP();
}
// Invalid if then nodes or graph changes.
private void calculateArrowsBackward(Node x, Node y, Graph graph) {
if (x == y) {
return;
}
if (!graph.isAdjacentTo(x, y)) {
return;
}
if (!knowledgeEmpty()) {
if (!getKnowledge().noEdgeRequired(x.getName(), y.getName())) {
return;
}
}
List<Node> naYX = getNaYX(x, y, graph);
clearArrow(x, y);
List<Node> _naYX = new ArrayList<Node>(naYX);
DepthChoiceGenerator gen = new DepthChoiceGenerator(_naYX.size(), _naYX.size());
int[] choice;
while ((choice = gen.next()) != null) {
List<Node> H = GraphUtils.asList(choice, _naYX);
if (!knowledgeEmpty()) {
if (!validSetByKnowledge(y, H)) {
continue;
}
}
double bump = deleteEval(x, y, H, naYX, graph);
if (bump > 0.0) {
Arrow arrow = new Arrow(bump, x, y, H, naYX);
sortedArrows.add(arrow);
addLookupArrow(x, y, arrow);
}
}
}
private void initializeArrowsBackward(Graph graph) {
sortedArrows.clear();
lookupArrows.clear();
for (Edge edge : graph.getEdges()) {
Node x = edge.getNode1();
Node y = edge.getNode2();
if (!knowledgeEmpty()) {
if (!getKnowledge().noEdgeRequired(x.getName(), y.getName())) {
continue;
}
}
if (Edges.isDirectedEdge(edge)) {
calculateArrowsBackward(x, y, graph);
} else {
calculateArrowsBackward(x, y, graph);
calculateArrowsBackward(y, x, graph);
}
}
}
private void addRequiredEdges(Graph graph) {
if (true) return;
if (knowledgeEmpty()) return;
for (Iterator<KnowledgeEdge> it = getKnowledge().requiredEdgesIterator(); it.hasNext(); ) {
KnowledgeEdge next = it.next();
Node nodeA = graph.getNode(next.getFrom());
Node nodeB = graph.getNode(next.getTo());
if (!graph.isAncestorOf(nodeB, nodeA)) {
graph.removeEdges(nodeA, nodeB);
graph.addDirectedEdge(nodeA, nodeB);
TetradLogger.getInstance()
.log("insertedEdges", "Adding edge by knowledge: " + graph.getEdge(nodeA, nodeB));
}
}
for (Edge edge : graph.getEdges()) {
final String A = edge.getNode1().getName();
final String B = edge.getNode2().getName();
if (knowledge.isForbidden(A, B)) {
Node nodeA = edge.getNode1();
Node nodeB = edge.getNode2();
if (nodeA != null
&& nodeB != null
&& graph.isAdjacentTo(nodeA, nodeB)
&& !graph.isChildOf(nodeA, nodeB)) {
if (!graph.isAncestorOf(nodeA, nodeB)) {
graph.removeEdges(nodeA, nodeB);
graph.addDirectedEdge(nodeB, nodeA);
TetradLogger.getInstance()
.log("insertedEdges", "Adding edge by knowledge: " + graph.getEdge(nodeB, nodeA));
}
}
if (!graph.isChildOf(nodeA, nodeB)
&& getKnowledge().isForbidden(nodeA.getName(), nodeB.getName())) {
if (!graph.isAncestorOf(nodeA, nodeB)) {
graph.removeEdges(nodeA, nodeB);
graph.addDirectedEdge(nodeB, nodeA);
TetradLogger.getInstance()
.log("insertedEdges", "Adding edge by knowledge: " + graph.getEdge(nodeB, nodeA));
}
}
} else if (knowledge.isForbidden(B, A)) {
Node nodeA = edge.getNode2();
Node nodeB = edge.getNode1();
if (nodeA != null
&& nodeB != null
&& graph.isAdjacentTo(nodeA, nodeB)
&& !graph.isChildOf(nodeA, nodeB)) {
if (!graph.isAncestorOf(nodeA, nodeB)) {
graph.removeEdges(nodeA, nodeB);
graph.addDirectedEdge(nodeB, nodeA);
TetradLogger.getInstance()
.log("insertedEdges", "Adding edge by knowledge: " + graph.getEdge(nodeB, nodeA));
}
}
if (!graph.isChildOf(nodeA, nodeB)
&& getKnowledge().isForbidden(nodeA.getName(), nodeB.getName())) {
if (!graph.isAncestorOf(nodeA, nodeB)) {
graph.removeEdges(nodeA, nodeB);
graph.addDirectedEdge(nodeB, nodeA);
TetradLogger.getInstance()
.log("insertedEdges", "Adding edge by knowledge: " + graph.getEdge(nodeB, nodeA));
}
}
}
}
}
/** Do an actual deletion (Definition 13 from Chickering, 2002). */
private void delete(Node x, Node y, List<Node> subset, Graph graph, double bump) {
Edge trueEdge = null;
if (trueGraph != null) {
Node _x = trueGraph.getNode(x.getName());
Node _y = trueGraph.getNode(y.getName());
trueEdge = trueGraph.getEdge(_x, _y);
}
if (log && verbose) {
Edge oldEdge = graph.getEdge(x, y);
String label = trueGraph != null && trueEdge != null ? "*" : "";
TetradLogger.getInstance()
.log(
"deletedEdges",
(graph.getNumEdges() - 1)
+ ". DELETE "
+ oldEdge
+ " "
+ subset
+ " ("
+ bump
+ ") "
+ label);
out.println(
(graph.getNumEdges() - 1)
+ ". DELETE "
+ oldEdge
+ " "
+ subset
+ " ("
+ bump
+ ") "
+ label);
} else {
int numEdges = graph.getNumEdges() - 1;
if (numEdges % 50 == 0) out.println(numEdges);
}
graph.removeEdge(x, y);
for (Node h : subset) {
Edge oldEdge = graph.getEdge(y, h);
graph.removeEdge(y, h);
graph.addDirectedEdge(y, h);
if (log) {
TetradLogger.getInstance()
.log("directedEdges", "--- Directing " + oldEdge + " to " + graph.getEdge(y, h));
}
if (verbose) {
out.println("--- Directing " + oldEdge + " to " + graph.getEdge(y, h));
}
if (Edges.isUndirectedEdge(graph.getEdge(x, h))) {
if (!graph.isAdjacentTo(x, h))
throw new IllegalArgumentException("Not adjacent: " + x + ", " + h);
oldEdge = graph.getEdge(x, h);
graph.removeEdge(x, h);
graph.addDirectedEdge(x, h);
if (log) {
TetradLogger.getInstance()
.log("directedEdges", "--- Directing " + oldEdge + " to " + graph.getEdge(x, h));
}
if (verbose) {
out.println("--- Directing " + oldEdge + " to " + graph.getEdge(x, h));
}
}
}
}
// serial.
private void insert(Node x, Node y, List<Node> t, Graph graph, double bump) {
if (graph.isAdjacentTo(x, y)) {
return; // The initial graph may already have put this edge in the graph.
// throw new IllegalArgumentException(x + " and " + y + " are already adjacent in
// the graph.");
}
Edge trueEdge = null;
if (trueGraph != null) {
Node _x = trueGraph.getNode(x.getName());
Node _y = trueGraph.getNode(y.getName());
trueEdge = trueGraph.getEdge(_x, _y);
}
graph.addDirectedEdge(x, y);
if (log) {
String label = trueGraph != null && trueEdge != null ? "*" : "";
TetradLogger.getInstance()
.log(
"insertedEdges",
graph.getNumEdges()
+ ". INSERT "
+ graph.getEdge(x, y)
+ " "
+ t
+ " "
+ bump
+ " "
+ label);
} else {
int numEdges = graph.getNumEdges() - 1;
if (verbose) {
if (numEdges % 50 == 0) out.println(numEdges);
}
}
if (verbose) {
String label = trueGraph != null && trueEdge != null ? "*" : "";
out.println(
graph.getNumEdges()
+ ". INSERT "
+ graph.getEdge(x, y)
+ " "
+ t
+ " "
+ bump
+ " "
+ label);
} else {
int numEdges = graph.getNumEdges() - 1;
if (verbose) {
if (numEdges % 50 == 0) out.println(numEdges);
}
}
for (Node _t : t) {
Edge oldEdge = graph.getEdge(_t, y);
if (oldEdge == null) throw new IllegalArgumentException("Not adjacent: " + _t + ", " + y);
graph.removeEdge(_t, y);
graph.addDirectedEdge(_t, y);
if (log && verbose) {
TetradLogger.getInstance()
.log("directedEdges", "--- Directing " + oldEdge + " to " + graph.getEdge(_t, y));
out.println("--- Directing " + oldEdge + " to " + graph.getEdge(_t, y));
}
}
}
private double andersonDarlingPASquareStarB(Node node, List<Node> parents) {
List<Double> _residuals = new ArrayList<Double>();
Node _target = node;
List<Node> _regressors = parents;
Node target = getVariable(variables, _target.getName());
List<Node> regressors = new ArrayList<Node>();
for (Node _regressor : _regressors) {
Node variable = getVariable(variables, _regressor.getName());
regressors.add(variable);
}
double sum = 0.0;
DATASET:
for (int m = 0; m < dataSets.size(); m++) {
RegressionResult result = regressions.get(m).regress(target, regressors);
TetradVector residualsSingleDataset = result.getResiduals();
for (int h = 0; h < residualsSingleDataset.size(); h++) {
if (Double.isNaN(residualsSingleDataset.get(h))) {
continue DATASET;
}
}
DoubleArrayList _residualsSingleDataset =
new DoubleArrayList(residualsSingleDataset.toArray());
double mean = Descriptive.mean(_residualsSingleDataset);
double std =
Descriptive.standardDeviation(
Descriptive.variance(
_residualsSingleDataset.size(),
Descriptive.sum(_residualsSingleDataset),
Descriptive.sumOfSquares(_residualsSingleDataset)));
// By centering the individual residual columns, all moments of the mixture become weighted
// averages of the moments
// of the individual columns.
// http://en.wikipedia.org/wiki/Mixture_distribution#Finite_and_countable_mixtures
for (int i2 = 0; i2 < _residualsSingleDataset.size(); i2++) {
// _residualsSingleDataset.set(i2, (_residualsSingleDataset.get(i2) - mean) /
// std);
// _residualsSingleDataset.set(i2, (_residualsSingleDataset.get(i2)) / std);
if (isMeanCenterResiduals()) {
_residualsSingleDataset.set(i2, (_residualsSingleDataset.get(i2) - mean));
}
}
double[] _f = new double[_residuals.size()];
for (int k = 0; k < _residuals.size(); k++) {
_f[k] = _residuals.get(k);
}
sum += new AndersonDarlingTest(_f).getASquaredStar();
}
return sum / dataSets.size();
}
private double localScoreB(Node node, List<Node> parents) {
double score = 0.0;
double maxScore = Double.NEGATIVE_INFINITY;
Node _target = node;
List<Node> _regressors = parents;
Node target = getVariable(variables, _target.getName());
List<Node> regressors = new ArrayList<Node>();
for (Node _regressor : _regressors) {
Node variable = getVariable(variables, _regressor.getName());
regressors.add(variable);
}
DATASET:
for (int m = 0; m < dataSets.size(); m++) {
RegressionResult result = regressions.get(m).regress(target, regressors);
TetradVector residualsSingleDataset = result.getResiduals();
DoubleArrayList _residualsSingleDataset =
new DoubleArrayList(residualsSingleDataset.toArray());
for (int h = 0; h < residualsSingleDataset.size(); h++) {
if (Double.isNaN(residualsSingleDataset.get(h))) {
continue DATASET;
}
}
double mean = Descriptive.mean(_residualsSingleDataset);
double std =
Descriptive.standardDeviation(
Descriptive.variance(
_residualsSingleDataset.size(),
Descriptive.sum(_residualsSingleDataset),
Descriptive.sumOfSquares(_residualsSingleDataset)));
for (int i2 = 0; i2 < _residualsSingleDataset.size(); i2++) {
_residualsSingleDataset.set(i2, (_residualsSingleDataset.get(i2) - mean) / std);
}
double[] _f = new double[_residualsSingleDataset.size()];
for (int k = 0; k < _residualsSingleDataset.size(); k++) {
_f[k] = _residualsSingleDataset.get(k);
}
DoubleArrayList f = new DoubleArrayList(_f);
for (int k = 0; k < f.size(); k++) {
f.set(k, Math.abs(f.get(k)));
}
double _mean = Descriptive.mean(f);
double diff = _mean - Math.sqrt(2.0 / Math.PI);
score += diff * diff;
if (score > maxScore) {
maxScore = score;
}
}
double avg = score / dataSets.size();
return avg;
}
/**
* Calculates the error variance for the given error node, given all of the coefficient values in
* the model.
*
* @param error An error term in the model--i.e. a variable with NodeType.ERROR.
* @return The value of the error variance, or Double.NaN is the value is undefined.
*/
private double calculateErrorVarianceFromParams(Node error) {
error = semGraph.getNode(error.getName());
Node child = semGraph.getChildren(error).get(0);
List<Node> parents = semGraph.getParents(child);
double otherVariance = 0;
for (Node parent : parents) {
if (parent == error) continue;
double coef = getEdgeCoefficient(parent, child);
otherVariance += coef * coef;
}
if (parents.size() >= 2) {
ChoiceGenerator gen = new ChoiceGenerator(parents.size(), 2);
int[] indices;
while ((indices = gen.next()) != null) {
Node node1 = parents.get(indices[0]);
Node node2 = parents.get(indices[1]);
double coef1, coef2;
if (node1.getNodeType() != NodeType.ERROR) {
coef1 = getEdgeCoefficient(node1, child);
} else {
coef1 = 1;
}
if (node2.getNodeType() != NodeType.ERROR) {
coef2 = getEdgeCoefficient(node2, child);
} else {
coef2 = 1;
}
List<List<Node>> treks = GraphUtils.treksIncludingBidirected(semGraph, node1, node2);
double cov = 0.0;
for (List<Node> trek : treks) {
double product = 1.0;
for (int i = 1; i < trek.size(); i++) {
Node _node1 = trek.get(i - 1);
Node _node2 = trek.get(i);
Edge edge = semGraph.getEdge(_node1, _node2);
double factor;
if (Edges.isBidirectedEdge(edge)) {
factor = edgeParameters.get(edge);
} else if (!edgeParameters.containsKey(edge)) {
factor = 1;
} else if (semGraph.isParentOf(_node1, _node2)) {
factor = getEdgeCoefficient(_node1, _node2);
} else {
factor = getEdgeCoefficient(_node2, _node1);
}
product *= factor;
}
cov += product;
}
otherVariance += 2 * coef1 * coef2 * cov;
}
}
return 1.0 - otherVariance <= 0 ? Double.NaN : 1.0 - otherVariance;
}