PaperlayerView exact revision on arXiv

Source-linked AI summary

Deep Skew-t Mixture Models

Jinran Wu, You-Gan Wang, Geoffrey J. McLachlan

arXiv:2609.00773v1stat.MEstat.COstat.ML

TL;DR

High-dimensional clustering needs models that handle heavy tails and directional asymmetry without making covariance estimation prohibitively complex. The paper proposes DStMM, which propagates a shared scale through deep factor-analytic pathways to obtain exact GHST components and evaluates it through stochastic/Monte Carlo EM. DStMM matches the symmetric robust model without skewness and gains increasingly as asymmetry strengthens, while real-data applications support its clustering and architecture-selection benefits.

  • Problem

    High-dimensional clustering requires component models that accommodate heavy tails and directional asymmetry while controlling covariance complexity.

  • Method

    DStMM propagates a shared inverse-gamma mixing variable through complete deep latent pathways, preserving conditional Gaussianity and exact GHST marginal components.

  • Results

    DStMM performs similarly to the symmetric robust model without skewness and yields increasing pathway-ARI gains as skewness strengthens, reaching 0.094 at κ = 2.

  • Takeaways & Limitations

    Exact pathway likelihoods support posterior path allocation and likelihood-based architecture comparison while skewness and heavy tails are propagated through the deep hierarchy.

  • Takeaways & Limitations

    With M = 1, stochastic-EM updates do not guarantee monotone observed-log-likelihood increase, and increasing depth requires enumerating complete latent paths.

Abstract

from arXiv · show

High-dimensional clustering is challenging when component distributions are both heavy-tailed and directionally asymmetric. We propose a deep skew-$t$ mixture model (DStMM), a hierarchical factor-analytic mixture based on the generalised-hyperbolic skew-$t$ normal mean--variance representation. A shared inverse-gamma mixing variable is propagated along each complete latent pathway, allowing heavy tails and directional asymmetry to be modelled jointly while preserving conditional Gaussianity. Each complete pathway therefore admits an exact GHST marginal representation. We formalise the reductions to symmetric deep $t$, Gaussian deep-mixture, and single-layer GHST factor-analytic models, discuss local non-identifiability and the implementation-level parameter-counting convention, and derive the conditional generalised inverse Gaussian law used for estimation. Estimation is carried out by a stochastic/Monte Carlo EM algorithm, with an explicit implementation-based parameter count for BIC architecture comparison. Simulation studies show that DStMM performs similarly to the symmetric robust model when skewness is absent but provides increasing gains as directional asymmetry becomes stronger, particularly under heavier tails; the same qualitative behaviour persists under smaller samples and unequal mixture proportions. Two real-data applications provide complementary evidence. On the UCI handwritten-digit benchmark, DStMM gives the strongest clustering performance under a common deep architecture, while on the Gas Sensor Array Drift data, DStMM improves on both deep Gaussian and deep $t$ alternatives and, under the implemented BIC criterion, selects a non-trivial second mixture layer. Together, these results support the value of propagating skewness and heavy-tail variation through a deep latent mixture while retaining an exact pathway-level likelihood.

1 Introduction

DStMM addresses high-dimensional clustering with component distributions that may be heavy-tailed and directionally asymmetric, while controlling covariance complexity through a deep factor-analytic mixture. Its shared pathway scale preserves conditional Gaussianity and gives every complete pathway an exact GHST observation-level representation.

  • Gaussian mixtures impose symmetric component geometry and can be sensitive to extreme observations, motivating models that accommodate heavy tails and directional asymmetry.
  • Deep Gaussian mixtures recursively compose factor-analytic transitions to combine dimension reduction, local parameter sharing, and flexible marginal structure.
  • DStMM propagates a common positive mixing variable through each complete latent pathway, jointly controlling variance inflation and directional mean displacement.
  • Conditional Gaussianity allows latent states to be marginalised exactly, so every complete pathway reduces to a multivariate GHST component and an exact observed-data likelihood.
  • The model reduces to Student-t RDMM under zero local skewness, to DGMM as degrees of freedom increase, and to a single-layer GHST factor-analytic mixture with one latent layer.
  • The paper derives pathway recursions, nested-model reductions, conditional GIG distributions, a stochastic/Monte Carlo EM algorithm, and an implementation-based BIC parameter count.

2 A deep pathway construction with GHST marginals

DStMM constructs a hierarchical mixture in which a shared inverse-gamma variable propagates through each latent pathway, jointly inducing heavy tails and directional skewness while preserving conditional Gaussianity. Each complete pathway collapses to an exact GHST component, supporting likelihood-based inference and reductions to familiar models.

  • Pathway construction: The hierarchy contains h latent mixture layers, with each layer having local components and complete paths formed by their component indicators.The term “deep” refers to recursively composed latent mixture transitions, not a deterministic neural-network architecture.
  • Pathway construction: A single inverse-gamma variable is reused across every transition on a selected path, controlling both dispersion and directional displacement at the observation level.The same path-specific mixing variable links variance inflation with the skewness-induced mean shift.
  • GHST pathway marginals: Proposition 1 shows that every fixed path has a Gaussian conditional marginal and an exact GHST marginal after integrating over the shared mixing variable.This closure property eliminates numerical integration over Gaussian latent states for pathway likelihood calculations.
  • GHST pathway marginals: The observed density is a finite mixture over complete paths, although posterior layer indicators generally become dependent after conditioning on observations.Prior path probabilities factorise across layers, whereas the joint pathway likelihood couples all path indices.
  • Estimation quantities: The conditional law of the shared scale is GIG, with analytic moments supporting posterior allocation and stable scale-related estimation updates.When the skewness-dependent term vanishes, the conditional law reaches an inverse-gamma limit; the symmetric case recovers Student-t down-weighting.
  • Nested-model reductions: The model reduces exactly to the symmetric Student-t RDMM when all local skewness parameters vanish, to DGMM as degrees of freedom increase, and to a GHST factor-analyser mixture when h = 1.These reductions position DStMM between Gaussian deep mixtures, symmetric heavy-tailed deep mixtures, and single-layer skew-t factor-analytic models.
  • Identifiability and parameterisation: Identifiability is handled separately at pathway and local levels through a pathway-separation assumption, an active skewness set, and an implementation-based loading parameter count.Factor-loading coordinates and intermediate latent coordinates remain non-unique, while richer active skewness allocations require additional care in parameter counting.

3 Estimation by stochastic/Monte Carlo EM

The DStMM is estimated with a stochastic or Monte Carlo EM scheme that collapses complete pathways, simulates shared-scale latent variables, and updates local parameters through weighted Gaussian regressions. Exact pathway enumeration creates the main computational burden, while architecture comparison uses an implementation-based BIC parameter count.

  • Latent-variable simulation: The augmented-data scheme conditionally simulates positive mixing variables and Gaussian latent states, reusing one scale draw throughout each pathway.Replacing the shared draw with layer-specific draws would define a different model and remove the exact pathway GHST marginal.
  • Algorithm: One simulated draw per observation–path pair yields a stochastic-EM approximation, whereas multiple draws produce a Monte Carlo EM approximation with less simulation noise.The degrees-of-freedom update can use analytic GIG moments and a safeguarded one-dimensional root finder.
  • M-step: Conditional on augmented variables, each transition becomes a heteroscedastic Gaussian regression, enabling weighted updates of intercepts, loadings, and active-layer skewness.Monte Carlo cross-products pool information over paths using the same local component, while covariance and mixing-proportion updates follow separately.
  • E-step: Pathway probabilities are computed from the observed GHST mixture, after collapsing each complete path to its induced parameters.The implementation evaluates these probabilities on the log scale and uses log-sum-exp normalization.
  • Fitting safeguards: With M = 1, monotone observed-log-likelihood increase is not guaranteed, so fitting monitors moving-window likelihood summaries and retains covariance lower bounds.Multiple or warm starts are recommended for the multimodal deep-mixture likelihood, with the retained fit selected by observed-data likelihood rather than external labels.
  • Computational scaling: Exact posterior allocation enumerates all complete paths, making increasing depth computationally limiting despite layer-wise parameter sharing.Direct dense calculations scale as O(|S|p^3) for collapsed covariance factorization and O(n|S|p^2) for likelihood evaluation; pathwise computations can be parallelised.
  • Model selection: Architecture comparison uses BIC(M) = −2ℓ(bΘ_M) + dcode(M) log n, with dcode(M) taken from the fitting implementation rather than a gauge-quotiented model-manifold dimension.The convention includes software-selected degrees-of-freedom contributions and applies consistently across DStMM candidate architectures.

4 Numerical simulations

The simulations test whether DStMM preserves symmetric-model performance while improving hierarchical recovery as skewness and tail heaviness increase. Across sample-size and proportion robustness checks, gains appear chiefly in complete-pathway allocation rather than first-layer clustering.

  • Simulation design: The experiments vary directional skewness, tail weight, sample size, and layer proportions to distinguish Gaussian, symmetric robust, and skewed heavy-tailed regimes.The main designs use a two-layer hierarchy and compare first-layer with complete-pathway clustering.
  • Experiment 1: Under symmetry, RDMM and DStMM have nearly identical complete-pathway ARI, 0.767 and 0.770, while both exceed 0.95 ARI at the first layer.This supports DStMM as an empirically effective extension of the symmetric robust model.
  • Experiment 1: DStMM’s pathway-ARI gain over RDMM increases from 0.013 at κ = 0.5 to 0.094 at κ = 2, while first-layer ARIs remain similar at 0.941 and 0.947.The improvement is therefore concentrated in deeper pathway allocation, where the hierarchy distinguishes substructure.
  • Experiment 1: DGMM’s mean pathway ARI falls from 0.489 at κ = 0 to 0.132 at κ = 2, whereas RDMM retains shared-scale robustness and DStMM adds scale-dependent directional displacement.The comparison separates the contribution of heavy-tail modelling from the additional skewness mechanism.
  • Experiment 2: The tail–skewness interaction strengthens DStMM’s advantage: at (ν, κ) = (5, 2), pathway ARI is 0.644 for DStMM, 0.529 for RDMM, and 0.042 for DGMM.At (15, 2), the corresponding ARIs are 0.775, 0.747, and 0.553, showing a smaller gap with lighter tails.
  • Robustness checks: The findings persist at n = 500, where DStMM’s pathway ARI is 0.625 versus 0.543 for RDMM at κ = 2, and under unequal layer proportions, where the advantage remains.At n = 500, the DStMM–RDMM gain is 0.082, close to 0.094 in the n = 1000 design.

5 Real applications

The real-data applications show that DStMM consistently outperforms deep Gaussian and symmetric deep t alternatives, with component-specific skewness contributing beyond heavy-tail robustness. It also supports richer architecture selection on the gas-sensor data.

  • 5.1 Case 1: Handwritten Digits: DStMM achieved the strongest controlled-architecture clustering on handwritten digits, with ARI 0.7325 and MR 0.2195.RDMM achieved ARI 0.7232 and MR 0.2240, while DGMM achieved ARI 0.6840 and MR 0.2355.
  • 5.1 Case 1: Handwritten Digits: 0.0485 higher ARI and 0.0160 lower MR than DGMM were observed for DStMM on handwritten digits.Relative to RDMM, DStMM increased ARI from 0.7232 to 0.7325 and reduced MR from 0.2240 to 0.2195.
  • 5.1 Case 1: Handwritten Digits: Component-specific skewness estimates displayed heterogeneous signed feature patterns across the ten handwritten-digit first-layer components.This supports estimating local skewness separately rather than imposing a common direction.
  • 5.2 Case 2: Gas Sensor Array Drift Data: The gas-sensor variables showed pronounced non-Gaussianity, with median absolute skewness 2.377 and median excess kurtosis 10.317.Absolute skewness exceeded one for 96.1% of variables, and excess kurtosis exceeded three for 82.0%.
  • 5.2 Case 2: Gas Sensor Array Drift Data: DStMM selected (K1, K2, r1, r2) = (6, 3, 16, 4) for gas sensors, attaining ARI 0.2989 and MR 0.4717.Selected RDMM attained ARI 0.2014 and MR 0.5357, while selected DGMM attained ARI 0.1803 and MR 0.5438.
  • 5.2 Case 2: Gas Sensor Array Drift Data: On the gas-sensor benchmark, DStMM’s ARI 0.2989 exceeded the listed published reference values, including SNN-DPC’s 0.7186 only in the separate handwritten-digit context.The published-reference comparisons are presented as benchmark context rather than strictly controlled pairwise comparisons.

6 Discussion

The discussion emphasizes a shared pathway scale that yields exact GHST observation-level components while preserving conditional Gaussian transitions. It also identifies computational, multimodality, Monte Carlo, and parameter-counting limitations.

  • 6 Discussion: A common positive scale simultaneously controls variance inflation and directional displacement in each exact GHST pathway distribution.Conditional Gaussianity permits exact marginalisation of latent states and an exact pathway likelihood.
  • 6 Discussion: The model nests Gaussian deep mixtures, symmetric heavy-tailed deep mixtures, and single-layer skew-t factor-analytic mixtures.These reductions connect the proposed construction to familiar model regimes.
  • 6 Discussion: Exact pathway likelihoods support posterior path allocation and likelihood-based architecture comparison without Monte Carlo approximation at the observation level.Stochastic simulation remains part of the local latent-hierarchy updates.
  • 6 Discussion: The implementation-based BIC follows diagonal-specific-covariance and rotation-counting conventions that define the reported architecture comparisons.The baseline active set A = {1} fixes where directional asymmetry is allocated in the reported fits.
  • 6 Discussion: Complete-path enumeration grows as |S| = Q_l K_l, so computational cost can increase rapidly with depth despite shared local parameters.The observed-data likelihood is multimodal, and M = 1 updates are subject to Monte Carlo variability; multiple starts and fit selection are therefore important.

A.1 Experiment 2 parameter recovery

As degrees of freedom increase, the latent scale becomes more concentrated, and parameter-recovery plots reflect reduced scale-mixing variation while clustering remains strong.

  • A.1 Experiment 2 parameter recovery: As ν increases, latent scale W becomes more concentrated and recovery patterns continue to track the model’s mean–variance structure.Across lighter-tailed settings, clustering performance remains strong.

A.2 First-layer robustness

Robustness checks show that RDMM and DStMM retain similar first-layer performance under reduced samples and unequal mixture proportions, with skewness helping mainly at complete-pathway resolution.

  • A.2 First-layer robustness: RDMM and DStMM retain similar first-layer performance across reduced-sample and unbalanced-mixture robustness designs.The principal benefit of modelling skewness appears in complete-pathway discrimination rather than the coarse first-layer partition.

A.3 Computational profile and fitting stability

The simulation runs show that DStMM is numerically stable across settings, while its richer Monte Carlo fitting procedure requires more time than DGMM and RDMM.

  • 1.38, 3.38, and 6.97 seconds are the Experiment 1 median elapsed times per successful fit for DGMM, RDMM, and DStMM, respectively.The corresponding Experiment 2 medians are 1.33, 3.11, and 6.58 seconds.
  • DStMM remains within single-digit seconds in these simulation settings despite its Monte Carlo E-step and additional skewness updates.
  • All 5000 Experiment 1 main fits succeed for DGMM, RDMM, and DStMM.
  • 9997, 10000, and 9999 of 10,000 Experiment 1 robustness fits succeed for DGMM, RDMM, and DStMM, respectively.
  • 9000 of 9000 Experiment 2 fits succeed for both RDMM and DStMM, compared with 8988 for DGMM.
Loading 2609.00773v1…