/**
* Compute relative variable importance for GBM model.
*
* <p>See (45), (35) formulas in Friedman: Greedy Function Approximation: A Gradient boosting
* machine. Algo used here can be used for computation individual importance of features per
* output class.
*/
@Override
protected VarImp doVarImpCalc(
GBMModel model, DTree[] ktrees, int tid, Frame validationFrame, boolean scale) {
assert model.ntrees() - 1 == tid
: "varimp computation expect model with already serialized trees: tid=" + tid;
// Iterates over k-tree
for (DTree t : ktrees) { // Iterate over trees
if (t != null) {
for (int n = 0; n < t.len() - t.leaves; n++)
if (t.node(n) instanceof DecidedNode) { // it is split node
Split split = t.decided(n)._split;
_improvPerVar[split._col] += split.improvement(); // least squares improvement
}
}
}
// Compute variable importance for all trees in model
float[] varimp = new float[model.nfeatures()];
int ntreesTotal = model.ntrees() * model.nclasses();
int maxVar = 0;
for (int var = 0; var < _improvPerVar.length; var++) {
varimp[var] = _improvPerVar[var] / ntreesTotal;
if (varimp[var] > varimp[maxVar]) maxVar = var;
}
// GBM scale varimp to scale 0..100
if (scale) {
float maxVal = varimp[maxVar];
for (int var = 0; var < varimp.length; var++) varimp[var] /= maxVal;
}
return new VarImp(varimp);
}
@Override
public void map(Chunk[] chks) {
_gss = new double[_nclass][];
_rss = new double[_nclass][];
// For all tree/klasses
for (int k = 0; k < _nclass; k++) {
final DTree tree = _trees[k];
final int leaf = _leafs[k];
if (tree == null) continue; // Empty class is ignored
// A leaf-biased array of all active Tree leaves.
final double gs[] = _gss[k] = new double[tree._len - leaf];
final double rs[] = _rss[k] = new double[tree._len - leaf];
final Chunk nids = chk_nids(chks, k); // Node-ids for this tree/class
final Chunk ress = chk_work(chks, k); // Residuals for this tree/class
// If we have all constant responses, then we do not split even the
// root and the residuals should be zero.
if (tree.root() instanceof LeafNode) continue;
for (int row = 0; row < nids._len; row++) { // For all rows
int nid = (int) nids.at80(row); // Get Node to decide from
if (nid < 0) continue; // Missing response
if (tree.node(nid) instanceof UndecidedNode) // If we bottomed out the tree
nid = tree.node(nid)._pid; // Then take parent's decision
DecidedNode dn = tree.decided(nid); // Must have a decision point
if (dn._split._col == -1) // Unable to decide?
dn = tree.decided(nid = dn._pid); // Then take parent's decision
int leafnid = dn.ns(chks, row); // Decide down to a leafnode
assert leaf <= leafnid && leafnid < tree._len;
assert tree.node(leafnid) instanceof LeafNode;
// Note: I can which leaf/region I end up in, but I do not care for
// the prediction presented by the tree. For GBM, we compute the
// sum-of-residuals (and sum/abs/mult residuals) for all rows in the
// leaf, and get our prediction from that.
nids.set0(row, leafnid);
assert !ress.isNA0(row);
double res = ress.at0(row);
double ares = Math.abs(res);
gs[leafnid - leaf] += _nclass > 1 ? ares * (1 - ares) : 1;
rs[leafnid - leaf] += res;
}
}
}
// --------------------------------------------------------------------------
// Build the next k-trees, which is trying to correct the residual error from
// the prior trees. From LSE2, page 387. Step 2b ii, iii.
private DTree[] buildNextKTrees(Frame fr) {
// We're going to build K (nclass) trees - each focused on correcting
// errors for a single class.
final DTree[] ktrees = new DTree[_nclass];
// Initial set of histograms. All trees; one leaf per tree (the root
// leaf); all columns
DHistogram hcs[][][] = new DHistogram[_nclass][1 /*just root leaf*/][_ncols];
for (int k = 0; k < _nclass; k++) {
// Initially setup as-if an empty-split had just happened
if (_distribution == null || _distribution[k] != 0) {
// The Boolean Optimization
// This optimization assumes the 2nd tree of a 2-class system is the
// inverse of the first. This is false for DRF (and true for GBM) -
// DRF picks a random different set of columns for the 2nd tree.
if (k == 1 && _nclass == 2) continue;
ktrees[k] = new DTree(fr._names, _ncols, (char) nbins, (char) _nclass, min_rows);
new GBMUndecidedNode(
ktrees[k],
-1,
DHistogram.initialHist(fr, _ncols, nbins, hcs[k][0], false)); // The "root" node
}
}
int[] leafs = new int[_nclass]; // Define a "working set" of leaf splits, from here to tree._len
// ----
// ESL2, page 387. Step 2b ii.
// One Big Loop till the ktrees are of proper depth.
// Adds a layer to the trees each pass.
int depth = 0;
for (; depth < max_depth; depth++) {
if (!Job.isRunning(self())) return null;
hcs = buildLayer(fr, ktrees, leafs, hcs, false, false);
// If we did not make any new splits, then the tree is split-to-death
if (hcs == null) break;
}
// Each tree bottomed-out in a DecidedNode; go 1 more level and insert
// LeafNodes to hold predictions.
for (int k = 0; k < _nclass; k++) {
DTree tree = ktrees[k];
if (tree == null) continue;
int leaf = leafs[k] = tree.len();
for (int nid = 0; nid < leaf; nid++) {
if (tree.node(nid) instanceof DecidedNode) {
DecidedNode dn = tree.decided(nid);
for (int i = 0; i < dn._nids.length; i++) {
int cnid = dn._nids[i];
if (cnid == -1
|| // Bottomed out (predictors or responses known constant)
tree.node(cnid) instanceof UndecidedNode
|| // Or chopped off for depth
(tree.node(cnid) instanceof DecidedNode
&& // Or not possible to split
((DecidedNode) tree.node(cnid))._split.col() == -1))
dn._nids[i] = new GBMLeafNode(tree, nid).nid(); // Mark a leaf here
}
// Handle the trivial non-splitting tree
if (nid == 0 && dn._split.col() == -1) new GBMLeafNode(tree, -1, 0);
}
}
} // -- k-trees are done
// ----
// ESL2, page 387. Step 2b iii. Compute the gammas, and store them back
// into the tree leaves. Includes learn_rate.
// gamma_i_k = (nclass-1)/nclass * (sum res_i / sum (|res_i|*(1-|res_i|)))
// For regression:
// gamma_i_k = sum res_i / count(res_i)
GammaPass gp = new GammaPass(ktrees, leafs).doAll(fr);
double m1class = _nclass > 1 ? (double) (_nclass - 1) / _nclass : 1.0; // K-1/K
for (int k = 0; k < _nclass; k++) {
final DTree tree = ktrees[k];
if (tree == null) continue;
for (int i = 0; i < tree._len - leafs[k]; i++) {
double g =
gp._gss[k][i] == 0 // Constant response?
? (gp._rss[k][i] == 0
? 0
: 1000) // Cap (exponential) learn, instead of dealing with Inf
: learn_rate * m1class * gp._rss[k][i] / gp._gss[k][i];
assert !Double.isNaN(g);
((LeafNode) tree.node(leafs[k] + i))._pred = g;
}
}
// ----
// ESL2, page 387. Step 2b iv. Cache the sum of all the trees, plus the
// new tree, in the 'tree' columns. Also, zap the NIDs for next pass.
// Tree <== f(Tree)
// Nids <== 0
new MRTask2() {
@Override
public void map(Chunk chks[]) {
// For all tree/klasses
for (int k = 0; k < _nclass; k++) {
final DTree tree = ktrees[k];
if (tree == null) continue;
final Chunk nids = chk_nids(chks, k);
final Chunk ct = chk_tree(chks, k);
for (int row = 0; row < nids._len; row++) {
int nid = (int) nids.at80(row);
if (nid < 0) continue;
ct.set0(row, (float) (ct.at0(row) + ((LeafNode) tree.node(nid))._pred));
nids.set0(row, 0);
}
}
}
}.doAll(fr);
// Collect leaves stats
for (int i = 0; i < ktrees.length; i++)
if (ktrees[i] != null) ktrees[i].leaves = ktrees[i].len() - leafs[i];
// DEBUG: Print the generated K trees
// printGenerateTrees(ktrees);
return ktrees;
}