Hierarchical Random Forest Models

• Hierarchical random forest models are extensions of standard ensembles that integrate hierarchical and multiscale decompositions to capture nested data structures.
• They leverage multi-stage pipelines, spatially recursive trees, and mixed-effects frameworks to address complex dependencies and improve prediction efficiency.
• Empirical applications in image segmentation, crystal structure prediction, and fairness-aware modeling demonstrate enhanced accuracy, macro-F1, and AUPR performance.

Hierarchical random forest models generalize the standard random forest framework to data with hierarchical, multiscale, or multi-resolution structure. By leveraging hierarchical decompositions either in feature construction, output structure modeling, or grouped data representations, these methods provide statistically and computationally efficient solutions for complex structured prediction problems across domains such as statistical inference, graph-based learning, fairness-aware modeling, and scientific image analysis.

1. Principle and Taxonomy of Hierarchical Random Forests

Hierarchical random forest models form an umbrella term encompassing a range of architectures that integrate tree-based ensembling into hierarchical problem decompositions. The underlying hierarchy may manifest as:

• Multi-stage model pipelines where the output from one random forest serves as input to another, capturing nested or sequential dependencies (e.g., crystal system → space group → lattice parameters in material science (Gleason et al., 2024)).
• Tree or pyramid structures mirroring data composition, such as multiresolution image patches in vision, where each tree node specializes to a particular spatial or feature granularity (Fallah, 2023).
• Models for grouped or longitudinal data (mixed effects), where random forests estimate nonparametric “fixed effect” components in the presence of group-level random effects (Bergonzoli et al., 2024).
• Hierarchical ensemble logic for fairness or latent attribute imputation, implementing two-layer forests to handle protected class proxies without explicit use (Li, 2021).
• Exploitation of intrinsic data geometry, such as recursive splits by horospheres in hyperbolic space to match tree-like or hierarchical representations (Doorenbos et al., 2023).

This taxonomy reflects a design flexibility to adapt random forest methodology to problem hierarchies defined by the data-generating process, output structure, or underlying geometry.

2. Model Architectures

The formal construction of hierarchical random forest models depends strongly on the application context. Prominent architectures include:

1. Hierarchical Multi-stage Forests: Decomposition of a global prediction task into a sequence or hierarchy of nested sub-tasks, each solved by a distinct (possibly task-specialized) random forest model. For example, the crystal structure prediction pipeline (Gleason et al., 2024) stacks a crystal-system classifier RF, per-system space-group classifier RFs, and per-system lattice-parameter regression RFs in a 3-level cascade. Each stage passes its prediction to condition the subsequent stage, with final inference obtained by hierarchical aggregation.
2. Spatially Recursive or Multiresolution Forests: Each tree is constructed to mirror the hierarchical decomposition of data (e.g., volumetric image patches). At each layer, trees process features at the corresponding resolution, recursively refining predictions on finer scales. The Hierarchical Quadratic Random Forest (Fallah, 2023) uses this technique for multichannel medical image segmentation, building trees that descend from coarse patches to voxels, passing context-aware features through the hierarchy.
3. Hierarchical Random Forests for Grouped Data: Random forests are embedded within a mixed-effects statistical framework, modeling hierarchical dependence between observations and grouping variables. The Ordinal Mixed-Effects Random Forest (OMERF) (Bergonzoli et al., 2024) replaces the linear fixed-effect of a cumulative link mixed model with a nonparametric forest, alternating between forest estimation and random effect inference via an EM-like scheme.
4. Proxy-aware Fairness via Hierarchical Forests: Two-layer architectures where a bottom-level forest predicts a latent protected attribute from proxies, followed by a top-level forest for the main prediction using both original and imputed features (Li, 2021). This approach aims to mitigate bias and enable "unaware fairness" under regulatory constraints.
5. Geometrically Hierarchical Forests: For data embedded in non-Euclidean or hierarchically structured spaces, hierarchical models exploit the geometry directly. Hyperbolic Random Forests (HoroRF) (Doorenbos et al., 2023) use horospheres—level sets determined by the Busemann function—as split surfaces in the Poincaré ball, reflecting the exponential scale of tree-like data.

3. Splitting Criteria and Node Models in Hierarchy

Hierarchical random forests generalize conventional splitting criteria and node models to align with the hierarchical task. Notable innovations include:

Approach	Splitting Mechanism	Notable Characteristics
Hyperbolic Random Forests (Doorenbos et al., 2023)	Horospherical splits via large-margin classifiers	Splits reflect hyperbolic geometry, addresses hierarchical class grouping via LCA.
Hierarchical Quadratic RF (Fallah, 2023)	Penalized multiclass linear discriminants on squared features	Quadratic decision boundaries in original feature space, group-Lasso feature selection.
OMERF (Bergonzoli et al., 2024)	Least squares on pseudo-responses, alternated with group fitting	Mixed-effect estimation, ordinal responses, EM-like alternation.
Fairness-aware HRF (Li, 2021)	Standard Gini/variance, two-level split sequence	Two-stage, explicit modeling of latent-protected class via proxy RF.
Crystal Structure RF (Gleason et al., 2024)	Gini impurity (classification), MSE (regression)	Pipeline: main label → sub-label → continuous, each specialized RF.

In hyperbolic forests, each split is a horosphere $\pi_{w,b}$ defined by an ideal boundary point $w$ (direction) and offset $b$, with candidate splits generated by large-margin optimization (HoroSVM), and information gain computed as in standard trees (Doorenbos et al., 2023). For multi-class hierarchies, class grouping is facilitated by hierarchical clustering in hyperbolic embedding space using lowest common ancestor (LCA) similarity.

The quadratic random forest model (Fallah, 2023) computes class-specific discriminant directions in a squared feature space, with node models estimated via group-Lasso regularized multiclass sparse discriminant analysis (MSDA); this produces sparse, node-specific supports and quadratic boundaries.

OMERF (Bergonzoli et al., 2024) uses the classic random forest split machinery, but the target variable at each split is a pseudo-response updated at every EM iteration to account for the latest random effect estimates, supporting inference at multiple hierarchical levels.

4. Training, Inference, and Aggregation

Training and prediction in hierarchical random forest models adhere to protocol-specific routines built atop recursive or staged random forest construction:

• Recursive Training Within Trees:

At each decision node, candidate splits and partitions are generated in alignment with the model’s structural priors (e.g., horospheres, quadratic projections, proxy imputation). Once stopping criteria are met (e.g., minimum samples, purity), leaves record empirical class probabilities or distributional estimates (Doorenbos et al., 2023, Fallah, 2023).

• Parallel and Staged Training:

Multi-stage models fit each RF per stage, conditioning child RFs on parent predictions. For instance, crystal structure prediction executes three distinct RFs per inference, with per-stage training data stratified by parent RF output (Gleason et al., 2024). Parallelization is commonly applied at the tree and stage level.

• EM-like Iterative Schemes:

OMERF alternates between fitting forests for fixed effects using pseudo-responses and updating random effects and thresholds via cumulative link mixed models until convergence (Bergonzoli et al., 2024).

• Inference:

Prediction requires traversing each relevant tree in the ensemble. For multi-pattern or multi-instance aggregation (e.g., in diffraction analysis), per-instance predictions are weighted according to the confidence gap between first and second most voted labels, then combined via “difference aggregation” (Gleason et al., 2024). In OMERF, new observations’ predicted logits are obtained via final forest and group-level random effects.

5. Empirical Performance and Applications

Hierarchical random forest models have demonstrated empirical effectiveness in domains where hierarchical or group structure is essential. Key findings:

• Hyperbolic Random Forests:

Achieve superior macro-F1 and AUPR across both imbalanced class and multi-class node classification on WordNet, graph node classification, and hierarchical vision benchmarks using hyperbolic embeddings, outperforming both Euclidean forests and state-of-the-art hyperbolic single-split classifiers (Doorenbos et al., 2023).

• Hierarchical Quadratic Random Forest:

Outperforms standard univariate and oblique forests in medical image segmentation, with group-Lasso improving generalization and efficiency in high-dimensional multiresolution feature spaces (Fallah, 2023).

• OMERF:

Achieves near state-of-the-art in complex nonlinear data-generating processes and in real data (PISA 2022), revealing non-linear group-level effects. In linear-dominated tasks, conventional CLMMs still lead (Bergonzoli et al., 2024).

• Unaware Fairness HRF:

Prediction accuracy and interval coverage for criminal justice tasks match or exceed naive models with explicit protected class; achieves regulatory compliance by handling proxies only (Li, 2021).

• Crystal Structure Prediction:

Hierarchical pipeline RF achieves 79% accuracy on crystal-system identification when aggregating across multiple diffraction patterns. Space-group and lattice parameter predictions attain 70–90% accuracy and 0.01–0.5 Å median errors for high-symmetry systems, with uncertainty and feature importance quantifiable at all stages (Gleason et al., 2024).

6. Limitations and Extensions

Several consistent limitations and further directions recur:

• Dependence on proxy quality (in fairness HRFs), and the challenge of optimally selecting informative proxies when protected classes are not directly observed (Li, 2021).
• Potential underperformance with insufficient or poorly structured hierarchies, or when the model hierarchy mismatches data generation (e.g., in OMERF, simpler linear models may outperform when true effects are linear) (Bergonzoli et al., 2024).
• Computational demands, which scale with the depth and width of hierarchies, mitigated by bagging, parallelism, and feature selection strategies (Fallah, 2023).
• Generalization to deeper hierarchies or additional structural priors—such as longitudinal data, multiple protected attributes, more intricate latent variable models, or integration with fully Bayesian schemes—remain open areas (Doorenbos et al., 2023, Bergonzoli et al., 2024).

Hierarchical random forest models thus unify multi-level inference, geometry-aware partitioning, proxy-based reasoning, and compositional ensembling under a flexible algorithmic regime. Their continued development and empirical validation expand the applicability of random forest methodology to complex data structures beyond the reach of standard tree ensembles.