Abstract
Bayesian Networks (BNs) are powerful tools for reasoning under uncertainty but often function as isolated models. This paper proposes a four-layered framework to bridge these "islands" by mapping semantically similar variables across different networks. This approach allows probabilistic evidence to flow between models without requiring them to be identical or merged into a single, computationally expensive system.
Introduction
In the field of Artificial Intelligence, Bayesian Networks are typically deployed as standalone systems. However, this approach faces significant challenges regarding scalability and distributed knowledge. As a network expands to cover a larger problem scope, inference becomes computationally sluggish and maintenance becomes difficult. Furthermore, different experts often create BNs for specific sub-domains using varying terminology or distributions for similar concepts. Existing solutions, such as Multiply Sectioned Bayesian Networks (MSBNs), require variables to be identical to link networks. This paper introduces a more flexible mapping system capable of handling semantically similar but non-identical variables.
The Mapping Framework
The proposed framework organizes BN mapping into four conceptual layers designed to manage the complexity of cross-network communication:
- Pair-wise Probabilistic Relations: Identifying joint probabilities that represent how a variable in Network A relates to a variable in Network B.
- Pair-wise Variable Linkages: Establishing "channels" for propagation using a similarity relation $Rel(A, B)$.
- Valid BN Mapping: Selecting a consistent subset of linkages to ensure networks remain synchronized.
- Minimum Valid BN Mapping: A reduction process to find the smallest set of linkages necessary to maintain consistency, preventing redundant computations.
Inference within this framework utilizes Jeffrey’s Rule and soft evidence propagation. When an observation is made in a source network, it is treated as soft evidence for the destination network, updating beliefs based on the probabilistic similarity between the mapped variables.
Key Insights: Similarity and the Ripple Effect
A core innovation of this framework is the emphasis on similarity over identity. Unlike previous methods, this model allows for "fuzzy" connections. If two variables represent similar concepts—such as "Precipitation" in one network and "Raining" in another—they can be linked via a joint distribution even if their underlying probability spaces differ.
Furthermore, the framework leverages the internal structure of BNs to create a "ripple effect." Mapping a single pair of variables can automatically satisfy probabilistic relations for other variables through standard internal inference.
As illustrated in the figure above, establishing a linkage between "precipitation" (Node A) and "Raining" (Node C) correctly updates the beliefs for dependent variables like "I use an umbrella" (Node D) across the network boundary. This "mapping reduction" makes the system significantly more scalable by minimizing the number of required cross-network links.
Results and Future Work
Validation using synthetic Bayesian Networks confirmed that the mapping reduction process effectively maintains consistency while reducing computational overhead. The framework proves more expressive than models like MSBN or AEBN by handling non-identical variables and diverse probability spaces.
Future research will focus on formalizing automated reduction rules and testing the framework's scalability in massive real-world systems. Additionally, the authors aim to apply this framework to ontology mapping, potentially bridging different semantic web ontologies using probabilistic logic.
Summary updated