Minimum Localized Bayesian Networks

• Minimum Localized Bayesian Networks are a framework that constructs minimal subnetworks guaranteeing exact inference for selected variable sets.
• They leverage directed convexity and iterative absorption algorithms to preserve all marginal and conditional distributions accurately.
• Empirical studies show substantial node reduction and up to 50× faster inference compared to traditional Bayesian network methods.

Minimum Localized Bayesian Networks (MLBNs) represent a rigorous framework for minimizing the complexity of Bayesian networks while preserving the full inferential power for selected variables. MLBNs leverage structural reductions, scoring-based selection, and context-specific parameterizations to yield models that are both parsimonious and inferentially sound. This article details the foundational principles, algorithmic construction, theoretical guarantees, and empirical characteristics of MLBNs, with an emphasis on the interplay between structural dimension reduction and localized parameter learning.

1. Formal Definition and Theoretical Foundations

Given a Bayesian network $\mathcal{B} = (G, \mathbb{P}(G))$, where $G = (V, E)$ is a directed acyclic graph (DAG), the objective of minimum localization is, for a specified query set $S \subseteq V$, to construct the smallest subnetwork that exactly preserves all marginal and conditional distributions involving only $S$ (and potentially an additional set of query variables $Q$ disjoint from $S$).

Minimum Localized Bayesian Network (MLBN):

A subset $H \subseteq V$ is the MLBN-node set for $S$ if:

• (i) The family of marginals of $\mathbb{P}(G)$ on $X_H$ coincides with the family of distributions factorizing according to the induced subgraph $G_H$;
• (ii) No strict subset $H' \supseteq S$ of $H$ satisfies this property.

This translates to MLBNs being the minimal subnetworks onto which the original network can be "collapsed" without altering the relevant inferences for variables in $S$ (Heng et al., 13 Jan 2026).

2. Directed Convexity, Inducing Paths, and the Directed Convex Hull

The critical combinatorial concept underpinning MLBNs is directed convexity ("d-convexity"):

• A subgraph $G_H$ is d-convex if no inducing path of $H$ exists—that is, there is no simple path $u \to \cdots \to v$ (with $u, v \in H$ and nonadjacent) such that all intermediate colliders are in $H$ and are ancestors of $u$ or $v$, and all non-colliders are outside $H$.
• The directed convex hull $\mathrm{CH}_G(S)$ is defined as the intersection of all d-convex supersets of $S$ in $G$.

The central equivalence theorem formalizes that: Theorem: Under faithfulness of $G$ to $\mathbb{P}(G)$, the MLBN-node set $H^* = \mathrm{CH}_G(S)$, and the MLBN for $S$ is exactly the DAG induced by the directed convex hull of $S$.

Thus, the MLBN construction is entirely characterized by the unique minimal d-convex superset of $S$ (Heng et al., 13 Jan 2026).

3. Algorithms for Extraction and Complexity

The construction of MLBNs reduces to computing the directed convex hull. The principal algorithm is CMDSA (Close Minimal D-Separator Absorption):

1. Initialize $H \leftarrow S$.
2. Iteratively, for each pair $u \neq v$ in $H$ that are nonadjacent but not d-separated by $H \setminus \{u, v\}$:
   • Identify minimal d-separators $S_u$ and $S_v$ (based on the Markov boundaries and Bayes-ball traversal).
   • Absorb these into $H$ ($H \leftarrow H \cup S_u \cup S_v$).
3. Terminate when no inducing pair remains.

Each absorption strictly enlarges $H$ until d-convexity is achieved. The overall complexity is $O(|V| \cdot (|V| + |E|))$ per query set $S$. The resulting $H = \mathrm{CH}_G(S)$ comprises the node set for the MLBN (Heng et al., 13 Jan 2026).

4. Inference Consistency and Faithfulness Guarantees

For any disjoint sets $Q$ and $S$,

$P_G(Q \mid S) = P_{\mathcal{B}_H}(Q \mid S),$

where $\mathcal{B}_H$ is the MLBN induced by $\mathrm{CH}_G(S)$. All d-separation relations among subsets of $S \cup Q$ are preserved, and the factorization of $P(x_{S \cup Q})$ remains identical in the MLBN and the original BN. This property ensures exact preservation of all query answers regarding $S$ after reduction. The faithfulness assumption is required for the correctness of this equivalence, as the characterization hinges on graphical criteria (Heng et al., 13 Jan 2026).

5. Empirical Performance: Dimension Reduction and Inference Speed

MLBNs via the directed convex hull demonstrate substantial node reduction and inference acceleration:

• Dimension Reduction Capability (DRC): Measured as $1 - |\mathrm{CH}_G(S)|/|V|$, DRC on real benchmarks includes Alarm (62.7%), Hepar2 (75.1%), Andes (70.9%), Diabetes (55.4%), Link (95.4%), Munin2 (98.7%).
• Inference Times: Constrained Variable Elimination (Con+VE) on the d-convex hull yields order-of-magnitude speedups over traditional variable elimination (VE) and belief propagation (BP), especially evident on large, sparse networks (e.g., Link: VE=305ms, Con+VE=7.43ms).
• Parameter Learning: KL-divergence between ground-truth marginals and re-learned submodels on the hull is negligible for moderate-to-large data ($\leq 0.006$ at $N=5000$ for random sparse networks) (Heng et al., 13 Jan 2026).

6. Local Structure and Parameter Minimality

Beyond structural dimension reduction, the notion of "minimum localized" also encompasses context-specific parameter minimization.

• Local Structure Models: Decision trees, default tables, and decision graphs are used for context-specific parameterization of CPTs, reducing parameter complexity from exponential in parent set size to a function of the number of distinct local contexts (Friedman et al., 2013, Chickering et al., 2013).
• Parameter Counting: The total parameter count becomes $P(G, S) = \sum_{i=1}^n q_i^S (r_i - 1)$, with $q_i^S$ the number of distinct structural or context-specific partitions.
• Learning Algorithms: Global search over graph structures is performed in tandem with greedy or pruned search over local CPT structures, jointly minimizing description length (MDL) or maximizing marginal likelihood (BDeu) (Friedman et al., 2013, Chickering et al., 2013).
• Log-linear Models and Causal Independence: First-Order Models (FOMs), restricting local conditionals to logit models with only additive (no interaction) parent effects, achieve further parameter reduction (from $O((T_Y-1)\prod_i T_i)$ to $O((T_Y-1)(1+\sum_i (T_i-1)))$) (Neil et al., 2013).

These approaches preserve the "minimum localized" property in the sense of parameter economy for a fixed structure or for hull-induced subnetworks.

7. Limitations, Extensions, and Related Methodologies

Limitations:

• MLBN reduction becomes less effective when $|S|$ is large or $G$ is highly connected, as $\mathrm{CH}_G(S) \approx V$.
• Each new query set $S$ requires recomputation of the directed convex hull.

Possible Extensions:

• Incorporation of hard evidence by including observed variables in $S$ and pruning barren nodes.
• Dynamic maintenance of the convex hull under incremental changes of $S$.
• Use of the d-convex hull for scalable structure learning or model decomposition.

Relation to Local Structure Discovery and Localized Learning:

Approaches such as SLL (Score-based Local Learning) focus on discovering the Markov blanket or local subgraphs for individual variables via local scoring and symmetry correction, often serving as preliminary steps for global assembly or local-to-global heuristics (Niinimaki et al., 2012). MDL decomposition and search frameworks (Lam et al., 2013) exploit local node-based scoring for scalable and interpretable structure discovery, with mechanisms for local updates and expert constraint integration.

Summary Table: Key MLBN Construction and Performance Results

Aspect	MLBN via Hull (Heng et al., 13 Jan 2026)	Local Parameter Learning (Friedman et al., 2013)
Node Reduction	d-convex hull extraction	Irrelevant for parameterization
Parameter Reduction	Induced subgraph only	Context-specific CPTs (decision trees, graphs)
Inference Consistency	Exact on $S$	Dependent on parameter model fit
Algorithmic Cost	$O(|V|(|V|+|E|))$	Typically modest over standard CPT
Empirical Gains	55–99% node reduction; 10–50$\times$ speedup	20–80% fewer parameters; faster convergence
Preservation Criteria	Faithfulness, d-convexity	Score equivalence, local structure

Minimum Localized Bayesian Networks offer a principled means of reducing both network structure and parametrization to the minimum required for answering all queries about a set $S$, with rigorous guarantees for marginal and conditional consistency, algorithmic tractability, and empirical efficacy in large-scale graphical models (Heng et al., 13 Jan 2026, Friedman et al., 2013, Chickering et al., 2013, Niinimaki et al., 2012, Neil et al., 2013, Lam et al., 2013).