Technical Summary: Laplace Variational Inference for Dirichlet Process Mixtures of Marked Poisson Point Processes

Problem Statement

The paper addresses the challenge of clustering replicated marked Poisson point processes (MPPP) observed on continuous domains. In many applications (e.g., biology, cosmology, sports analytics), data consists of multiple subjects, each generating a finite collection of events with associated attributes (marks). The inferential goal is to identify latent groups of subjects that share similar event intensity structures and mark distributions.

Existing approaches face significant limitations:

1. Discretization: Many methods require gridding the continuous domain or thinning the process, which introduces approximation errors and computational instability, especially in high dimensions or irregular geometries.
2. Model Rigidity: Previous clustering methods often rely on parametric families or fail to jointly infer the number of clusters and continuous intensity surfaces.
3. Inference Complexity: Standard log-link models (Log-Gaussian Cox Processes) suffer from intractable likelihood integrals, while alternative links (like the squared link) introduce non-conjugacy and optimization pathologies (sign ambiguity and nodal lines).

The authors aim to develop a Bayesian nonparametric model that jointly infers latent cluster assignments, the number of clusters, and continuous mark-specific intensity surfaces without discretizing the domain, while providing a scalable and theoretically sound inference algorithm.

Methodology

1. Model Formulation: DPM-MPPP

The authors propose a Dirichlet Process Mixture of Marked Poisson Point Processes (DPM-MPPP).

• Data Structure: For each subject $i$, observations consist of points $Y_i$ in a domain $B$ and binary marks $m_i \in \{0, 1\}$. An offset $T_i$ (e.g., time, area) accounts for varying exposure levels.
• Clustering Mechanism: Subjects are assigned to clusters $z_i$ via a Dirichlet Process (DP) with concentration parameter $\alpha$ and base measure $G_0$. The DP allows for nonparametric inference of the number of clusters.
• Intensity Modeling: Within cluster $k$, the intensity for mark $m$ is modeled as:
  $$\lambda_{km}(y) = h(W_{km}(y)) = (B(y)^\top \theta_{km})^2$$
  where $B(y)$ is a vector of non-negative basis functions (e.g., B-splines), $\theta_{km}$ are coefficients, and $h(x)=x^2$ is the squared link function.
• Key Feature: The model uses an offset $T_i$ to separate overall event volume from spatial and mark-specific structure. This ensures clustering is driven by spatial preferences and mark composition rather than total event counts.
• Mark Dependence: The model induces a spatially varying mark probability surface $p_k(y) = \lambda_{k1}(y) / (\lambda_{k0}(y) + \lambda_{k1}(y))$, capturing dependencies between event occurrence and marks.

2. Inference Algorithm: Constrained Laplace Variational Bayes

To perform posterior inference, the authors develop a Variational Bayes (VB) algorithm.

• Challenge: The squared link function yields a tractable likelihood integral (avoiding the need for grid discretization) but results in a non-conjugate posterior for the basis coefficients $\theta_{km}$. Standard coordinate ascent fails to provide closed-form updates.
• Solution: The authors introduce a constrained Laplace approximation for the non-conjugate block $\theta_{km}$.
  • Sign Ambiguity & Nodal Lines: The squared link is invariant under $\theta \to -\theta$, creating sign ambiguity. Furthermore, if $\theta^\top B(y)$ crosses zero, it creates "nodal lines" where the intensity vanishes, leading to unstable optimization and local modes.
  • Constrained Optimization: To resolve this, the mode-finding problem is restricted to a strictly positive chamber $D_\delta = \{ \theta : \|\theta\|_2 \le R, \inf_{y \in B} B(y)^\top \theta \ge \delta \}$.
  • Algorithm Steps:
    1. Update conjugate blocks (stick-breaking weights, cluster assignments, variance parameters) via closed-form coordinate ascent.
    2. Update non-conjugate coefficients $\theta_{km}$ by finding the mode of the variational objective within $D_\delta$ using constrained optimization (e.g., L-BFGS-B).
    3. Approximate the posterior covariance of $\theta_{km}$ using the inverse Hessian evaluated at the constrained mode.

3. Theoretical Guarantees

The paper provides rigorous theoretical support for the constrained Laplace approximation:

• Consistency: The population maximizer of the objective over the positive chamber is unique and lies strictly inside the feasible set $D_\delta$, away from the zero boundary.
• Separation: Sign-changing coefficients (which cause nodal lines) are shown to have strictly lower empirical objective values than the positive (or negative) chamber modes.
• Exponential Dominance: The exponential weight of sign-changing configurations is asymptotically negligible compared to the positive chamber. Thus, restricting optimization to the positive chamber does not discard asymptotically relevant mass.

Key Contributions

1. Novel Model: Introduction of the DPM-MPPP, the first variational Bayes method for clustering replicated marked Poisson point processes that jointly infers cluster structure, cluster count, and continuous mark-specific intensity surfaces.
2. Grid-Free Inference: Utilization of the squared link function to obtain tractable continuous-domain likelihood terms without gridding, thinning, or finite-element approximations.
3. Algorithmic Innovation: Development of a constrained Laplace approximation that resolves the sign ambiguity and nodal-line pathologies inherent in squared-link models, enabling stable, grid-free, and efficient high-dimensional inference.
4. Theoretical Rigor: Establishment of theoretical guarantees proving that the constrained optimization targets the correct global mode and that sign-changing regions are asymptotically negligible.

Experimental Results

The authors evaluated the method through synthetic experiments and a real-world application:

• Synthetic Experiments:

  • Tested on diverse scenarios including mark-swapped clusters, high numbers of clusters (up to 9), irregular intensity surfaces, and varying sample sizes (sparse to dense).
  • Performance: The proposed method accurately recovered cluster-specific intensity surfaces, subject assignments, and the number of clusters. It demonstrated superior clustering purity compared to baseline feature-based methods (K-means and Gaussian Mixture Models on binned or kernel-smoothed data), particularly in reduced-sample regimes where feature summaries became noisy.
  • Robustness: The method remained stable even with significantly reduced event counts, showing only mild over-clustering (splitting small groups) rather than mixing distinct clusters.

• Real-World Application (NBA Shot Charts):

  • Applied to 219,527 shot attempts from 566 NBA players (2024–2025 season), modeling shot locations and make/miss outcomes.
  • Findings: The model identified 16 active clusters representing distinct player archetypes. It successfully resolved fine-grained heterogeneity within broad roles (e.g., distinguishing between different types of scorers or big men based on spatial shot selection and location-specific success probabilities).
  • Interpretability: The fitted surfaces provided interpretable insights into player strategies, separating spatial preference from shooting efficiency.

Significance and Claims

The paper claims that the proposed framework offers a principled, scalable, and theoretically grounded approach to clustering marked point processes on continuous domains. By combining the computational advantages of the squared link with a constrained Laplace approximation, the method overcomes the trade-offs between tractability and model flexibility that plague existing approaches. The authors emphasize that their method:

• Avoids the "curse of dimensionality" associated with grid-based methods.
• Provides a unified framework for modeling event intensity and mark distributions without discarding the continuous nature of the data.
• Delivers robust performance in sparse data regimes where traditional feature-based clustering fails.

The work is presented as a significant step forward in Bayesian nonparametric inference for spatial point processes, with immediate applicability to fields requiring the analysis of replicated event data with complex mark structures.