---
title: Explainable Tree-Ensemble Pipelines
url: https://www.emergentmind.com/topics/explainable-tree-ensemble-pipelines
type: topic
---

# Explainable Tree-Ensemble Pipelines

Shape Analysis Using Fractal Dimension: A Curvature-Based Approach

Fractal dimension methods provide quantitative measures of shape complexity and scaling behavior. In shape analysis, these methods capture the self-similarity and roughness of planar curves or boundaries, offering robust descriptors that are both informative and, in some designs, invariant under common geometric transformations. The curvature-based fractal dimension technique is a multiscale approach that integrates the classical notion of fractal dimension with a curvature scale-space framework, yielding multidimensional descriptors that enhance discrimination power and stability in pattern recognition tasks [1201.3118].

## 1. Theoretical Framework and Motivation

The fractal dimension $D$ is classically defined for geometric objects by the scaling of a covering number or measure as the measurement scale varies:
\[
D = \lim_{\epsilon \to 0} \frac{\log N(\epsilon)}{\log(1/\epsilon)}
\]
where $N(\epsilon)$ is the number of covering elements of linear size $\epsilon$. For planar contours, $N(\epsilon)$ may be replaced by quantities (such as curve length or summed curvature) that exhibit power-law behavior with scale: $L(\epsilon) \propto \epsilon^{-D}$. For empirical, pseudo-fractal contours, $D$ is estimated by performing linear regression on the log–log relationship between such measures and scale. The fractal dimension thus quantifies the degree to which a curve fills space: $D\approx1$ for smooth curves, $D\to2$ for highly convoluted or rough curves.

## 2. Curvature Scale-Space Formalism

Let $\mathbf{C}(s) = (x(s), y(s))$ denote a parameterized, closed planar contour. The curvature scale-space (CSS) technique applies Gaussian smoothing at each scale $\sigma$, producing smoothed coordinates:
\[
x_{\sigma}(s) = (g_{\sigma} * x)(s), \quad y_{\sigma}(s) = (g_{\sigma} * y)(s)
\]
where $g_{\sigma}(t) = (1/\sigma\sqrt{2\pi}) \exp(-t^2/(2\sigma^2))$ and $*$ denotes convolution. The curvature at scale $\sigma$ is given by:
\[
k_{\sigma}(s) = \frac{x_{\sigma}'(s) y_{\sigma}''(s) - y_{\sigma}'(s)x_{\sigma}''(s)}{(x_{\sigma}'(s)^2 + y_{\sigma}'(s)^2)^{3/2}}
\]
Numerical derivatives are computed using finite differences on the sampled contour.

## 3. Multiscale Fractal Dimension Descriptors

At each scale $\sigma_i$, compute the absolute summed curvature:
\[
S(\sigma_i) = \sum_{s} |k_{\sigma_i}(s)|
\]
The cumulative sum mimics dilation operations in traditional fractal estimation:
\[
A(\sigma_i) = \sum_{j=0}^i S(\sigma_j)\, \Delta \sigma
\]
Assuming a power law $A(\sigma) \propto \sigma^D$, the global fractal dimension $D$ is the slope of the log–log plot:
\[
\log A(\sigma) \approx D\, \log \sigma + \mathrm{const}
\]
For increased descriptive power, the method evaluates the local slope (multiscale fractal dimension):
\[
d(\sigma_i) = \frac{\log A(\sigma_{i+1}) - \log A(\sigma_{i})}{\log \sigma_{i+1} - \log \sigma_{i}}
\]
Sampling $d(\sigma_i)$ over a prescribed $\sigma$-range yields a multiscale fractal dimension vector, serving as a high-dimensional shape signature.

## 4. Descriptor Extraction Pipeline

The full extraction can be formalized as follows:

```python
# Pseudocode for curvature-based multiscale fractal dimension extraction
Input: sampled contour C = {(x_n, y_n)} for n = 1..N; scale parameters σ_0 < ... < σ_M
Output: descriptor vector D[0..M-1]

for i in 0..M:
    convolve x, y with Gaussian kernel g_{σ_i} to obtain x_i, y_i
    compute first and second derivatives: x_i', x_i'', y_i', y_i''
    compute curvature k_i[n] = (x_i'[n]*y_i''[n] - y_i'[n]*x_i''[n]) / (x_i'[n]^2 + y_i'[n]^2)^{3/2}
    S[i] = sum_{n=1}^N |k_i[n]|
A[0] = S[0]
for i in 1..M:
    A[i] = A[i-1] + S[i]
X[i] = log σ_i; Y[i] = log A[i]
for i in 0..M-1:
    D[i] = (Y[i+1] - Y[i]) / (X[i+1] - X[i])
return D[0..M-1]
```

Typical descriptor length $M-1$ is chosen based on application-specific requirements.

## 5. Comparative Experimental Performance

The multiscale curvature-based descriptors were validated on a 1,100-class fish silhouette dataset (11,000 shapes with controlled rotations and scales), using descriptor length $50$ and linear discriminant analysis with cross-validation. Results demonstrate:

| Descriptor Type | Accuracy (%) |
|----------------|-------------|
| Curvature-FD   | 97.14       |
| Bouligand–Minkowski (no norm) | 14.30 |
| Bouligand–Minkowski (diam norm)| 80.06|

The curvature-based method exhibits robust invariance to rotation and scale, a property resulting from the transformation-invariant nature of summed curvature and the intrinsic filtering effect of Gaussian smoothing. Its multiscale feature vector yields strong discriminability, far surpassing classic dilation-based multiscale fractal descriptors [1201.3118].

## 6. Advantages and Limitations

**Advantages:**
- Intrinsic invariance to rotation and scale due to curvature-summation and global convolution
- Robustness to noise and contour perturbation via progressive Gaussian smoothing
- Efficient offline extraction and suitability for high-throughput applications
- Rich multiscale characterization of local and global complexity, surpassing single-number descriptors

**Limitations:**
- Requires careful tuning of scale range $[\sigma_{min}, \sigma_{max}]$ and the number of scales $M$ to match the object's sampling density and application
- Numerical curvature computation may be unstable for poorly sampled or highly noisy contours
- Assumes closed, smooth contours; open or highly irregular curves may necessitate preprocessing

## 7. Position in Shape Analysis and Pattern Recognition

The curvature-based multiscale fractal dimension method situates itself among shape analysis approaches as a robust tool for quantifying complexity, distinct from both classic global FD estimators and spectral or Minkowski-based multiscale descriptors. By embedding the fractal-dimension estimation into the curvature scale-space, the method unifies the geometric intuition of space-filling with a scale-adaptive sensitivity to shape features. In comparative studies, it proves both highly discriminative and more resilient to normalization uncertainties—a favorable property for large-scale pattern recognition systems and morphometric analysis [1201.3118].

Source: https://www.emergentmind.com/topics/explainable-tree-ensemble-pipelines