pruning! leaves pruned nodes without sample-indices, breaking predict-forest/test-forest

[masatoi]

pruning! turns a split node back into a leaf via delete-children!, but it does not restore the node's sample-indices. Since make-forest defaults to :remove-sample-indices? t, every node that was split already had its sample-indices set to nil, so the newly created leaves have none. Any subsequent predict-forest / test-forest then fails.

Reproduction

Self-contained, no dataset download needed (SBCL 2.6.7, master @ b4e5ec5):

(ql:quickload :cl-random-forest)

(defparameter *seed* 42)
(defun nextrand ()
  (setf *seed* (mod (+ (* 1103515245 *seed*) 12345) 2147483648))
  (/ (float *seed* 1.0) 2147483648.0))
(defun noise () (* 2.0 (- (+ (nextrand) (nextrand) (nextrand)) 1.5)))

(defun gen (n-datum n-dim n-class signal)
  (let ((dm (make-array (list n-datum n-dim) :element-type 'single-float))
        (tg (make-array n-datum :element-type 'fixnum)))
    (dotimes (i n-datum)
      (let ((c (mod i n-class)))
        (setf (aref tg i) c)
        (dotimes (j n-dim)
          (setf (aref dm i j) (+ (noise) (if (= (mod j n-class) c) signal 0.0))))))
    (values dm tg)))

(multiple-value-bind (x y) (gen 2000 20 5 1.0)
  (defparameter *x* x)
  (defparameter *y* y))

;; NOTE: :remove-sample-indices? defaults to T
(defparameter *forest*
  (clrf:make-forest 5 *x* *y* :n-tree 30 :bagging-ratio 0.2 :max-depth 8 :n-trial 15))
(defparameter *rd* (clrf:make-refine-dataset *forest* *x*))
(defparameter *learner* (clrf:make-refine-learner *forest*))
(clrf:train-refine-learner *learner* *rd* *y*)

(clrf:test-forest *forest* *x* *y* :quiet-p t)   ; => 85.1   (OK)

(clrf:pruning! *forest* *learner* 0.2)

(clrf:test-forest *forest* *x* *y* :quiet-p t)   ; => TYPE-ERROR

Result:

The value
  NIL
is not of type
  (SIMPLE-ARRAY FIXNUM)
from the function type declaration.

  0: CL-RANDOM-FOREST/SRC/RANDOM-FOREST::NODE-CLASS-DISTRIBUTION (src/random-forest.lisp:221)
  1: CL-RANDOM-FOREST/SRC/RANDOM-FOREST::CLASS-DISTRIBUTION-FOREST (src/random-forest.lisp:707)
  2: CL-RANDOM-FOREST/SRC/RANDOM-FOREST:PREDICT-FOREST (src/random-forest.lisp:730)
  3: CL-RANDOM-FOREST/SRC/RANDOM-FOREST:TEST-FOREST (src/random-forest.lisp:739)

Cause

set-best-children! nils out the parent's sample-indices when the tree was built with the default :remove-sample-indices? t (src/random-forest.lisp:414-415):

(when (dtree-remove-sample-indices? dtree)
  (setf (node-sample-indices node) nil))

delete-children! (src/random-forest.lisp:1212-1217) makes that node a leaf again but never puts sample-indices back:

(defun delete-children! (node)
  (setf (node-test-attribute node) nil
        (node-test-threshold node) nil
        (node-left-node node) nil
        (node-right-node node) nil)
  node)

node-class-distribution (src/random-forest.lisp:221-224) recomputes the leaf distribution from sample-indices on every prediction, so it receives nil.

Why it is easy to miss

The refine-learner path keeps working after pruning, because make-refine-vector only reads node-leaf-index. So make-refine-dataset -> train-refine-learner -> test-refine-learner all succeed, and the failure only shows up when predict-forest / test-forest is called on the pruned forest.

Regression forests have the same structure through node-regression-mean (src/random-forest.lisp:515).

Workaround

Build the forest with :remove-sample-indices? nil. Verified that the exact same pruning sequence then completes and test-forest works (71.2% before pruning -> 71.5% after, on a 100-tree forest).

Suggested fix

Either

• make delete-children! restore the node's sample-indices by concatenating the sample indices of the subtree leaves being discarded, or
• have pruning! assert that the forest was built with :remove-sample-indices? nil, and document the requirement (the same way src/feature-importance.lisp needs it for the OOB set).

The first keeps pruning! usable with the memory-saving default; the second is the minimal change.

Related: #15 (forest-n-leaf is stale after pruning!).