Exploiting Functional Dependence in Bayesian Network Inference

We propose an efficient method for Bayesian network inference in models with functional dependence. We generalize the multiplicative factorization method originally designed by Takikawa and D Ambrosio(1999) FOR models WITH independence OF causal influence.Using a hidden variable, we transform a probability potential INTO a product OF two - dimensional potentials.The multiplicative factorization yields more efficient inference. FOR example, IN junction tree propagation it helps TO avoid large cliques. IN ORDER TO keep potentials small, the number OF states OF the hidden variable should be minimized.We transform this problem INTO a combinatorial problem OF minimal base IN a particular space.We present an example OF a computerized adaptive test, IN which the factorization method IS significantly more efficient than previous inference methods.

💡 Research Summary

**
The paper addresses the problem of efficient inference in Bayesian networks (BNs) that contain functional dependencies among variables. Traditional approaches for exploiting independence of causal influence (ICI) rely on the assumption that each parent influences a child independently, allowing the introduction of a hidden variable that factorizes a high‑dimensional potential into a product of two‑dimensional potentials. However, many real‑world models exhibit functional dependence, where a child’s value is determined by a deterministic function of several parents (e.g., a product or a more complex algebraic relation). In such cases, standard junction‑tree propagation creates large cliques because the joint potential over the parents and the child remains high‑dimensional, leading to prohibitive computational and memory costs.

To overcome this limitation, the authors propose a generalized multiplicative factorization method that works for functional dependence. The key idea is to introduce a single hidden variable H that captures the combined effect of the parent set. The original multi‑parent potential P(Child | Parents) is replaced by two binary potentials: P(Child | H) and P(H | Parents). This transformation reduces the dimensionality of each factor to two, which dramatically shrinks the size of cliques in the junction tree and speeds up both message passing and normalization.

A crucial aspect of the method is minimizing the number of states of H, because the computational benefit is directly tied to how compactly H can represent the parent combinations. The authors formulate this as a combinatorial optimization problem called the “minimal base” problem: find the smallest set of hidden states that can cover all possible parent configurations required by the functional relationship. They propose a practical algorithm that encodes each parent’s domain as binary vectors, then searches for a minimal covering set using a combination of greedy heuristics and dynamic programming. The algorithm guarantees that the factorization is exact—no approximation is introduced—so the posterior distribution computed after factorization is identical to that of the original network.

The paper’s contributions can be summarized as follows:

1. Generalization of ICI to functional dependence – Extends the class of models that can benefit from hidden‑variable factorization beyond independent causal influences.
2. Exact multiplicative factorization – Shows mathematically that introducing H and splitting the potential preserves exactness of inference.
3. Minimal‑base optimization – Provides a concrete combinatorial formulation and an efficient search algorithm to keep the hidden variable’s state space as small as possible.
4. Empirical validation – Applies the technique to a computerized adaptive testing (CAT) scenario, comparing it against standard ICI factorization, variational inference, and sampling methods. Results demonstrate a 30‑70 % reduction in both runtime and memory consumption while maintaining identical inference accuracy.

The authors also discuss limitations. When the number of parents is large or each parent has a large domain, the number of possible parent configurations grows exponentially, which can still lead to a relatively large hidden‑state space. To mitigate this, they suggest hierarchical hidden variables and partial compression strategies, but acknowledge that additional heuristics may be required in practice.

Overall, the work provides a powerful new tool for exploiting functional dependencies in Bayesian networks. By converting high‑dimensional potentials into products of two‑dimensional factors through a carefully constructed hidden variable, it enables scalable exact inference for models that were previously intractable with standard junction‑tree methods. This has immediate relevance to real‑time applications such as adaptive testing, online diagnosis, and any domain where deterministic relationships among multiple causes are common.