Variable Splitting Binary Tree (VSBT)

• VSBT is a tree-based model that uses recursive binary splits for unsupervised data segmentation in both clustering and time series environments.
• It employs a deviance-based split selection mechanism and variable split locations to ensure transparent and optimal segmentation.
• In its Bayesian formulation, VSBT integrates context-tree priors and variational approximations to efficiently quantify uncertainty in tree structures and regime assignments.

The Variable Splitting Binary Tree (VSBT) is a family of interpretable, tree-based models for unsupervised data segmentation through recursive binary splits, adaptable both to clustering in multivariate settings and to change-point detection in time series. VSBT models operate by recursively partitioning the sample or temporal space using axis-parallel or interval splits, followed by systematic aggregation steps. The method offers transparent segmentations and statistically rigorous clustering or segmentation regimes, admitting both frequentist and Bayesian formulations. Core themes include deviance-based split selection and variable split locations, structural and probabilistic tree priors, and efficient pruning/agglomeration mechanisms (Fraiman et al., 2011, Nakahara et al., 22 Jan 2026).

1. Model Definition and Scope

VSBT is defined as a hierarchical, top–down splitting procedure. Given observations $X_i$ from an unknown distribution $P$, VSBT recursively builds a maximal binary tree, with each node corresponding to a subset of the data (for clustering) or a time interval (for time series). Splits are axis-parallel in the clustering regime (Fraiman et al., 2011) or specified by flexible, recursive logistic regression models for time segmentation (Nakahara et al., 22 Jan 2026). Terminal nodes, or leaves, define clusters (clustering) or AR/i.i.d. regimes (time series segmentation).

In the Bayesian time series context, internal nodes carry submodels parameterized by logistic regression coefficients $(\beta_{s,0}, \beta_{s,1})$, which select split positions within intervals. Each leaf is assigned a generative AR or i.i.d. submodel.

2. Splitting Criteria and Structural Mechanisms

Clustering Framework

The splitting criterion is the deviance functional $$R(t) = \alpha_t\,\mathrm{tr}[\mathrm{Cov}(X|X\in t)]$$, which is minimized by splitting a node $t$ into two subregions $t_l$ and $t_r$. The gain in deviance,

$$R(t) - R(t_l) - R(t_r) = \frac{\alpha_{t_l}\alpha_{t_r}}{\alpha_t} \|\mu_{t_l} - \mu_{t_r}\|^2,$$

is maximized over potential splits. Splits are axis-parallel: for variable $j$ and threshold $a$, the two subregions are $t_l = \{x: x(j) \leq a\} \cap t$, $t_r = \{x: x(j) > a\} \cap t$. Sampling-based analogues substitute empirical means and covariances (Fraiman et al., 2011).

Time Series Segmentation

The tree structure encodes interval partitioning through recursive logistic regression:

$$P(u_{t,d_s}=1 \mid \beta_s) = \sigma(\beta_{s,0} t + \beta_{s,1}),\quad \sigma(z) = \frac{1}{1+e^{-z}},$$

where $u_{t,d_s}$ denotes the path choice at depth $d_s$ for time $t$. The model allows split locations to be arbitrary within each interval, leading to compact trees, unlike fixed-split context tree models (Nakahara et al., 22 Jan 2026). Each leaf $s$ is assigned an AR model, $$x_t \sim \mathcal N(\tilde{x}_t^T \theta_{k(s)}, \tau_{k(s)}^{-1})$$.

3. Pruning, Joining, and Agglomeration Procedures

Pruning (Clustering)

Sibling leaves $(t_l,t_r)$ are merged if their empirical supports are sufficiently similar. Dissimilarity is measured by computing $\delta$-quantile-based "nearest neighbor distances" $\bar{d}_l^\delta$, $\bar{d}_r^\delta$ and defining $$d^\delta(t_l,t_r) = \max(\bar{d}_l^\delta, \bar{d}_r^\delta)$$. If $d^\delta$ is below a user-specified threshold $\epsilon$ (mindist), pairs are collapsed (Fraiman et al., 2011).

Global Joining

Final clusters or regimes are formed by joining any pair of leaves with sufficiently similar empirical representation. Joining proceeds either until a user-specified number of clusters $k$ is reached, or, if $k$ is unknown, until all pairwise distances exceed a threshold $\eta$ (often chosen as a low quantile post-pruning).

4. Bayesian Formulation and Context-Tree Priors

The Bayesian VSBT model for time segmentation employs context-tree weighting (CTW) priors for tree structures:

$$p(T) = \prod_{s \in \mathcal I_T} g_s \prod_{s \in \mathcal L_T} (1-g_s),$$

where $g_s$ is the split probability at node $s$. Regression coefficients $\beta_s$ have Gaussian priors, $\beta_s \sim \mathcal N(\eta_s, L_s^{-1})$. AR model assignments at leaves are categorical with Dirichlet-distributed parameters.

CTW recursion manages posterior computation over all tree structures:

$$g'_s = \frac{g_s \prod_{\text{ch}} \phi_{\text{ch}}}{(1-g_s)\sum_k \rho_{s,k} + g_s\prod_{\text{ch}}\phi_{\text{ch}}}$$

with detailed recursion for $\phi_s$ across nodes and leaf assignments (Nakahara et al., 22 Jan 2026).

5. Inference Algorithms and Complexity

Inference in the Bayesian VSBT uses mean-field variational approximation for the logistic factors (employing the Jaakkola–Jordan lower bound), CTW recursion for the tree posterior, and conjugate updates for AR parameters and assignments. Each iteration involves:

• Forward–backward updates for $q(\bm u)$ across the tree.
• Recursive computation of $\rho_{s,k}$, $\phi_s$, $g'_s$ for tree and regime assignment.
• Closed-form updates for AR parameters and Dirichlet assignments.
• Local Gaussian updates for $q(\beta_s)$ leveraging quadratic forms from the bound.
• Updates of local variational parameters $\xi_{s,t}$.

Overall complexity per iteration is $$O(n D_{\max} + |\mathcal I_{T_{\max}}| K^2 + D_{\max} K^3)$$ for AR updates and $O(n D_{\max})$ for logistic regression factors.

In clustering, maximal tree construction scales worst-case as $O(p n^2)$, with practical implementations achieving $O(p n \log n)$ via sorting and cumulative sum optimizations. Pruning and joining, exploiting quantile-based selection, stay computationally efficient for moderate $n$ and well-chosen $\delta$ (Fraiman et al., 2011).

6. Empirical Results and Illustrative Examples

Clustering Performance

VSBT achieves high interpretability and segmentation fidelity across simulated and real datasets (Fraiman et al., 2011). For four 2D Gaussian clusters, perfect allocation occurs for small $\sigma$; for high-dimensional (50D) Gaussian mixtures, VSBT matches or outperforms model-based clustering benchmarks. In the "European Jobs" dataset (25 countries × 9 sectors), VSBT identifies canonical splits aligned with economic-political groupings by agriculture and mining percentages.

Time Series Segmentation

Synthetic experiments with $n=75$ and two change points show VSBT recovers true segmentations at minimal tree depth (mean error $<1$ sample, mean depth $2.1$, parameter reduction $60\%$ compared to FSBT). Uncertainty quantification, via posterior probabilities for change-points, yields credible intervals $\sim 5$ samples wide. Replicated results demonstrate stability of tree depth and segmentation accuracy (Nakahara et al., 22 Jan 2026).

Method	Mean Error	Std Error	Mean Depth	# Params
FSBT	2.97	1.15	7.8	156
VSBT	0.92	0.31	2.1	48

Editor's term: FSBT refers to "fixed-split binary tree" segmentation.

7. Advantages, Limitations, and Hyperparameter Specification

VSBT achieves transparent, interpretable partitions and compact tree representations by learning split locations rather than relying on predetermined splits. Bayesian formulations explicitly quantify uncertainty in both split placement and regime assignment. Marginalization over trees via CTW is exact with respect to tree prior and posterior weights, and avoids sampling inefficiencies.

Limitations center on variational bias affecting posterior variance for logistic regression factors, and computational scaling with tree depth or number of AR models. Deep trees with many possible regimes incur higher cost.

Key hyperparameters include:

• $\mathrm{mindev} \in [0.7,0.9]$ for clustering split selection,
• $\mathrm{minsize}$ for minimal cluster size,
• $\delta \approx 0.2$ (robustness to outliers),
• $\epsilon$ (pruning threshold),
• $\eta$ (quantile for joining threshold).

A plausible implication is that VSBT offers a unifying, extensible platform for both multivariate clustering and time series segmentation, accommodating both frequentist deviance-based and Bayesian context-tree paradigms.