BDe Metric in Bayesian Networks

• BDe metric is a Bayesian network scoring function based on Dirichlet priors that ensures likelihood equivalence for Markov-equivalent structures.
• Its BDeu special case applies a uniform prior on local conditional probability tables but may bias results with degenerate or highly skewed data.
• The alternative GU metric employs a global uniform prior to robustly detect independence, though it remains computationally challenging for complex networks.

The BDe metric (Bayesian Dirichlet equivalent) is a foundational scoring function used for Bayesian network (BN) structure learning, grounded in marginal likelihood under Dirichlet priors that guarantee likelihood equivalence for Markov-equivalent structures. Its widely-used uniform special case, BDeu, distributes prior mass uniformly across local conditional probability tables (CPTs). Recent analysis has identified fundamental pathologies in BDeu's behavior with degenerate or highly skewed data, motivating the proposal of the Global Uniform (GU) metric, which places a single uniform prior on the set of joint distributions consistent with a BN structure, rather than independent local priors. While GU addresses key drawbacks of BDeu and K2, its implementation is computationally intractable for arbitrary BNs except in special cases. This article synthesizes the rigorous framework, mathematical properties, practical limitations, and research implications of the BDe family and the GU metric, using standard notations for discrete BN structure learning as described in (Kayaalp et al., 2012).

1. Formal Definition of the BDe Metric

Given a complete data set $D$ of $N$ instances of discrete variables $X_1,\ldots,X_n$, consider a BN structure $S$ encoding, for each variable $X_i$, a set of parents $\text{Pa}_i$ ($|\text{Pa}_i| = q_i$). Let $N_{ijk}$ denote the count in $D$ for $X_i=k$ given $\text{Pa}_i$ is in parent configuration $j$, and $N_{ij} = \sum_k N_{ijk}$. The BDe metric assigns Dirichlet priors to each local CPT:

• For node $i$ and parent configuration $j$, hyperparameters $\{\alpha_{ij1},...,\alpha_{ijr_i}\}$ (where $r_i$ is the number of values $X_i$ can take) are specified, with total prior mass $\alpha_{ij} = \sum_k \alpha_{ijk}$.
• The BDe marginal likelihood is

$$P(D|S) = \prod_{i=1}^n \prod_{j=1}^{q_i} \frac{\Gamma(\alpha_{ij})}{\Gamma(\alpha_{ij} + N_{ij})} \prod_{k=1}^{r_i} \frac{\Gamma(\alpha_{ijk} + N_{ijk})}{\Gamma(\alpha_{ijk})}$$

• Likelihood equivalence is guaranteed if $$\alpha_{i1} = ... = \alpha_{iq_i} \equiv \alpha_{i0}$$ for each node $i$. Structures encoding the same independencies then receive identical scores.

2. The BDeu Special Case

BDeu ("u" for uniform) imposes a single global equivalent sample size $\alpha_0$ and distributes it uniformly:

• $\alpha_{ijk} = \alpha_0/(r_i q_i)$ for all $i,j,k$
• $\alpha_{ij} = \alpha_0/q_i \equiv \alpha_{i0}$

The closed-form BDeu marginal likelihood becomes:

$$P(D|S) = \prod_{i=1}^n \prod_{j=1}^{q_i} \frac{\Gamma(\alpha_0/q_i)}{\Gamma(\alpha_0/q_i + N_{ij})} \prod_{k=1}^{r_i} \frac{\Gamma(\alpha_0/(r_i q_i) + N_{ijk})}{\Gamma(\alpha_0/(r_i q_i))}$$

BDeu is computationally efficient and guarantees likelihood equivalence. The prior is specified purely by $\alpha_0$; its interpretation as "equivalent sample size" guides practical selection, though sensitivity to this hyperparameter can impact model selection.

3. Pathologies of the BDeu Metric

BDeu exhibits two notable pathologies, each stemming from the uniform local Dirichlet prior:

• Degenerate-data bias toward dependence: When some variables have constant values (i.e., $N_{ijk}=0$ for some $k$ across all samples), BDeu disproportionately favors adding arcs ("dependent" structures), even if the correct generating structure is independent. This arises because unobserved configurations still receive nonzero prior mass, and extra arcs allow greater freedom to allocate this mass, inflating P(D|S) with increasing sample size.
• Sensitivity to skewed marginals: In cases where all $N_{ijk} > 0$ but distributions are highly skewed, the relative impact of the uniform prior versus empirical data may result in erratic scoring, sometimes leading to counterintuitive preference for dependence or independence as a function of $\alpha_0$, data size, and skewness.

In both phenomena, a fixed $\alpha_0$ applied locally induces a structure-dependent, non-uniform prior on the global joint parameter space, so that structures with different arcs correspond to fundamentally different implicit priors.

4. The Global Uniform Metric

The GU metric addresses the BDeu-induced prior inconsistency by specifying a uniform prior directly over the set $Q_S$ of joint distributions $\theta = P(X_1,...,X_n)$ that satisfy exactly the independence constraints of $S$. Formally,

• $f(\theta|S) = C$ if $\theta \in Q_S$; $0$ otherwise, with $C$ chosen for normalization ($\int_{Q_S} f(\theta|S)\,d\theta=1$).
• The GU score is the marginal likelihood

$$P(D|S) = \int_{Q_S} P(D|\theta)\,f(\theta|S)\,d\theta$$

Evaluating $P(D|S)$ under GU is tractable in two special cases:

1. Saturated Model (single clique):
   • Structure $S$ allows any joint distribution; $Q_S$ is the full simplex of dimension $r-1$.
   • The score becomes

   $$P(D|S) = \frac{\Gamma(r)}{\Gamma(r + \sum_i N_i)} \prod_{i=1}^r \Gamma(1 + N_i)$$

   where $r$ is the joint state space cardinality.

2. Independent Saturated Model (disconnected nodes):
   • $S$ factors into $m$ disconnected cliques.
   • The GU marginal likelihood factorizes:

   $$P(D|S) = \prod_{j=1}^m \frac{\Gamma(r_{(j)})}{\Gamma(r_{(j)} + \sum_i N_{(j),i})} \prod_{i=1}^{r_{(j)}} \Gamma(1 + N_{(j),i})$$

For arbitrary structures, the integral is over a high-dimensional, constraint-defined polytope and does not generally yield to closed-form evaluation or efficient computation.

5. Comparative Analysis: BDeu, K2, and GU

The following summarizes principal properties of these metrics:

Aspect	BDeu	K2	GU
Prior choice	Local Dirichlet $(\alpha_0/(r_i q_i),\ldots)$	Local Dirichlet $(1,\ldots,1)$	Uniform over global joint (no hyperparameter)
Parameter independence	Yes	Yes	No (only special factorable cases)
Modularity	Yes	Yes	No
Likelihood equivalence	Yes	No	Yes
Computational cost	$O(n r_i q_i)$, closed-form	$O(n r_i q_i)$, closed-form	Intractable except for cliques/disconnected
Hyperparameter	$\alpha_0$ user-chosen	None	None
Robustness (degenerate/skewed data)	Pathologies observable	Irreducible ordering/bias issues	Correctly favors independence, if computable

BDeu's tractability and equivalence properties make it the workhorse for structure learning, but its spurious dependence bias and sensitivity issues are documented. K2 offers a local uniform prior but lacks likelihood equivalence and can be influenced by variable ordering. GU sets a true uniform prior with respect to the global joint and is free from BDeu's identified pitfalls, but is generally infeasible computationally beyond small or structurally simple networks (Kayaalp et al., 2012).

6. Open Challenges and Research Directions

GU's main unresolved issue is computational intractability for general structures; only trivial (clique/disconnected) cases admit analytic solutions. Relevant open problems include:

• Derivation of closed-form or recursive solutions for additional classes of graphs (e.g., trees, polytrees).
• Development of approximation schemes for the GU-marginal likelihood, using methods such as Laplace approximation, variational bounds, or Monte Carlo integration (e.g., importance sampling over $Q_S$).
• Investigation of decomposable scoring functions that approximate GU while relaxing strict modularity.

Efficient, broadly-applicable implementations of GU or GU-inspired metrics remain a significant open area, with potential to bring likelihood-equivalent, globally-uniform priors to structure learning in practical settings.

7. Practical Implications and Summary

BDe and BDeu provide likelihood-equivalent and computationally efficient scoring for Bayesian-network structure learning under assumptions of local parameter independence and modularity. However, their implicit prior distributions may unduly favor dependence in degenerate data scenarios or misbehave under skewness. The GU metric restores uniformity over the exact set of distributions consistent with a structure and robustly identifies true independence, at the cost of general computational intractability. The search for tractable, accurate GU approximations or new structure priors that avoid BDeu's pathologies while retaining efficiency is an important avenue for future research (Kayaalp et al., 2012).