Forest of Model Trees

• Forest of model trees is an ensemble learning method where each tree partitions the input by data-adaptive hyperplanes and fits local linear models at the leaves.
• It employs convolutional regularization and smooth C¹ blending to produce continuously differentiable regressors, enhancing robustness against input perturbations.
• The training algorithm guarantees convergence by recursively fitting least-squares models with a tilt constraint, ensuring precise, low-variance predictions.

A forest of model trees is an ensemble learning approach in which each base learner is a model tree—specifically, a tree that partitions the input space by means of data-adaptive hyperplanes at internal nodes and fits local linear models at the leaves. Recent developments focus on application domains such as function approximation over high-dimensional images, where the method leverages down-sampling, convolutional regularization, and smooth C¹ blending to produce continuously differentiable regressors with provable convergence guarantees (Armstrong, 16 Nov 2025). These model tree forests stand in contrast to classical piecewise constant decision tree ensembles, and are conceptually related to, but distinct from, transformation forests that aggregate local conditional distribution models in a parametric framework (Hothorn et al., 2017).

1. Formal Definition and Structure

Let $d$ denote the (optionally down-sampled) dimensionality of the input, typically a vectorized image, and $H = \prod_{i=1}^d [0,w_i]$ denote an axis-aligned hyper-rectangle (HR). A model tree $T$ is a full binary tree with the following elements:

• Internal nodes $n$ store hyperplane (HP) functions $S_n: H \to \mathbb{R}$,
• Leaves $\ell$ store linear functions $F_\ell: H \to \mathbb{R}$.

Inputs $x \in H$ are routed down the tree according to the sign of $S_n(x)$. For node $n$ with center $H = \prod_{i=1}^d [0,w_i]$0 and least-squares fit coefficients $H = \prod_{i=1}^d [0,w_i]$1, the split function is

$H = \prod_{i=1}^d [0,w_i]$2

At a leaf $H = \prod_{i=1}^d [0,w_i]$3, the local regression is

$H = \prod_{i=1}^d [0,w_i]$4

where $H = \prod_{i=1}^d [0,w_i]$5 are least-squares coefficients, $H = \prod_{i=1}^d [0,w_i]$6 is the centroid of samples in the leaf, and $H = \prod_{i=1}^d [0,w_i]$7 is the average label.

A forest of such model trees comprises $H = \prod_{i=1}^d [0,w_i]$8 independently constructed trees, each providing both a prediction $H = \prod_{i=1}^d [0,w_i]$9 and a leaf-specific weight $T$0, combined into a weighted average output:

$T$1

2. Down-Sampling and Input Preprocessing

Prior to constructing model trees for image data, the images are down-sampled by partitioning the original grid into non-overlapping $T$2 blocks and representing each super-pixel by the average intensity in that block. This yields a lower-dimensional input vector $T$3 of length $T$4 for images of dimension $T$5. This dimensionality reduction not only accelerates least-squares fitting but also modifies the meaning of hyperplanes and regression coefficients, which now operate over super-pixels (Armstrong, 16 Nov 2025).

3. Convolutional Regularization of Hyperplanes

To impart robustness against localized distortions (e.g., minor translations or small deformations in images), the hyperplane coefficient grid $T$6 at each node is convolved with a spatial kernel $T$7. For a position $T$8 on the down-sampled grid, the convolved coefficients are:

$T$9

After reshaping $n$0 to match the spatial grid, this convolution is implemented prior to inference, meaning no additional runtime overhead at prediction. The resulting split tests $n$1 are computed using the convolved weights, enhancing generalization under slight input perturbations (Armstrong, 16 Nov 2025).

4. Forest Construction and Ensemble Prediction

A forest of model trees may be constructed by two principal methods:

• Training trees on independent bootstrap samples of the training data.
• Training a base tree, then perturbing the centers $n$2 of each node's HR by a small vector $n$3 and re-fitting the splits.

Each tree produces a weight $n$4 that is inherently smooth due to the application of smoothing kernels at all split nodes. The overall forest prediction $n$5 is their weighted average as specified above. The denominator is safeguarded against degenerate cases (i.e., $n$6) by including a "helper tree" with near-zero output and minimal weight (Armstrong, 16 Nov 2025).

Ensembling diverse trees in this manner ensures that the sharp discontinuities of individual trees are averaged out, reducing prediction variance.

5. Smooth Blending and Output Regularity

To overcome the inherent discontinuity of tree-based methods at split boundaries, model trees equip each node split with a $n$7-continuous smoothing kernel:

$n$8

with $n$9, where $S_n: H \to \mathbb{R}$0 is the margin width at node $S_n: H \to \mathbb{R}$1. The node-wise weight is $S_n: H \to \mathbb{R}$2, and the per-tree leaf weight is the product over the path to $S_n: H \to \mathbb{R}$3, $S_n: H \to \mathbb{R}$4. This smoothing produces a forest-level output $S_n: H \to \mathbb{R}$5 that is globally $S_n: H \to \mathbb{R}$6—i.e., continuously differentiable—provided the base function $S_n: H \to \mathbb{R}$7 is itself continuously differentiable and other technical criteria are met (Armstrong, 16 Nov 2025).

6. Training Algorithm and Convergence Guarantee

The training procedure recursively fits least-squares regressions at each node. If the RMSE of the fit is below a threshold $S_n: H \to \mathbb{R}$8, the block is made a leaf; otherwise, the most influential split axis is selected based on importances $S_n: H \to \mathbb{R}$9. A "tilt-constraint" with factor $\ell$0 enforces that only nearly-axis-aligned splits are permitted:

$\ell$1

where $\ell$2 is the axis of highest importance. If this constraint fails, low-importance coefficients are suppressed to ensure geometric shrinkage of the partition blocks. The algorithm guarantees convergence: for any $\ell$3 function $\ell$4 defined over $\ell$5, recursion halts in finite time, and in each leaf, the linear fit achieves RMSE at most $\ell$6 (Armstrong, 16 Nov 2025). No explicit regularization is required under idealized assumptions and with sufficiently dense sampling.

7. Comparison with Related Model Tree Forests

Model tree forests as described above are structurally different from transformation forests (Hothorn et al., 2017), although both aggregate tree-based predictors with locally adaptive models at the leaves. Whereas model trees partition feature space and fit piecewise linear regressors, transformation trees fit parametric transformation models at leaves that capture the entire conditional distribution. Transformation forests then aggregate these models via forest weights, yielding local maximum-likelihood estimates of the conditional law and facilitating prediction intervals and quantile regression. This suggests that the "forest of model trees" approach is particularly targeted at regression over high-dimensional structured domains (e.g., images), leveraging local linearity, convolutional structure, and explicit $\ell$7 smoothing, whereas transformation forests focus on local distribution estimation via adaptive likelihood aggregation.

---

Key References for this methodology:

• "Convolutional Model Trees" (Armstrong, 16 Nov 2025)
• "Transformation Forests" (Hothorn et al., 2017)