Skip to main content
Back to timeline
arXivSource publication:

Curve classifiers stop relying on one whole-curve decomposition: window-level components let GL-FRF beat a tuned full-curve forest on 16 of 20 UCR datasets

Synopsis

Functional random forests usually take one principal-component decomposition of the whole curve, which can dilute discriminative features confined to short intervals. GL-FRF stacks a tuned full-curve forest with one forest per window using penalised convex weights, and the full-curve forest remains a feasible solution; on 20 UCR datasets it improved on a tuned full-curve forest with the same number of trees on 16, with Brier score reductions up to 0.21, significant after Holm adjustment. WC-FRF trains a single forest on concatenated window scores, was more accurate on 13 real datasets, but degraded sharply in simulated settings where the signal required the full curve.

Source-provided article image: Global-Local and Window-Concatenated Functional Random Forests: Early and Late Fusion of Window-Specific Functional Principal Components
arXiv

Interpretation

The input representation for curve classification expands from one principal-component decomposition of the whole curve to the whole curve plus several equal-width windows, each smoothed and decomposed separately, with a choice of how to fuse them. Earlier functional random forests trained only on full-curve principal-component scores, so discriminative features confined to short intervals could be diluted by the variance structure of the whole curve while the signal location was unknown in advance. On 20 UCR datasets GL-FRF improved on a tuned full-curve forest with the same number of trees on 16, with a median Brier difference of 0.015 and the largest reductions on Trace (0.213) and FreezerRegularTrain (0.178); the difference remained significant after Holm adjustment.

GL-FRF treats the full-curve forest and each window forest as parallel experts and estimates convex combination weights from cross-fitted predictions, with the weights lying on the probability simplex. Stacking combines the predictions of several models; here two penalties shrink the total local weight toward the full-curve expert and the local weights toward equality. Weights are estimated from predictions of a three-fold cross-fitting layer inside the training set, and the number of windows and penalty strength are chosen by nested cross-validation; when local experts do not help the cross-fitted loss, the full-curve expert receives all the weight and the method reduces exactly to a tuned full-curve forest.

WC-FRF concatenates the principal-component scores of all windows into one feature vector and trains a single forest, letting trees choose windows at splits and express interactions between windows. Unlike GL-FRF, which trains per-window forests and then weights them, WC-FRF fuses before the forest, so it has no full-curve representation and no explicit fallback. On the 20 real datasets WC-FRF had a lower Brier score than GL-FRF on 13, with a median difference of 0.007, but the difference was not significant (adjusted 0.286); in the cross-window interaction simulation S5, WC-FRF scored 0.493, close to a constant prediction, against 0.079 for GL-FRF.

Perspective

The evaluation covers 12 simulated scenarios with known signal structure and 20 univariate equal-length UCR datasets with at least 12 training curves per class, and the comparison is limited to principal-component-based forests, random-interval forests and functional nearest neighbours.

The simulated scenarios did not reproduce the situations in which windows helped on real data, so when a window-specific representation outperforms a tuned global one remains to be characterised; the locality control also differs from GL-FRF in how curves are smoothed, so part of the difference may come from window-specific smoothing.