Fuzzy Knowledge Representation Based on Possibilistic and Necessary Bayesian Networks

Within the framework proposed in this paper, we address the issue of extending the certain networks to a fuzzy certain networks in order to cope with a vagueness and limitations of existing models for decision under imprecise and uncertain knowledge. This paper proposes a framework that combines two disciplines to exploit their own advantages in uncertain and imprecise knowledge representation problems. The framework proposed is a possibilistic logic based one in which Bayesian nodes and their properties are represented by local necessity-valued knowledge base. Data in properties are interpreted as set of valuated formulas. In our contribution possibilistic Bayesian networks have a qualitative part and a quantitative part, represented by local knowledge bases. The general idea is to study how a fusion of these two formalisms would permit representing compact way to solve efficiently problems for knowledge representation. We show how to apply possibility and necessity measures to the problem of knowledge representation with large scale data. On the other hand fuzzification of crisp certainty degrees to fuzzy variables improves the quality of the network and tends to bring smoothness and robustness in the network performance. The general aim is to provide a new approach for decision under uncertainty that combines three methodologies: Bayesian networks certainty distribution and fuzzy logic.

💡 Research Summary

The paper introduces a novel framework that merges Bayesian networks with fuzzy logic and possibilistic reasoning to handle both uncertainty and vagueness in knowledge representation. Traditional Bayesian networks excel at modeling stochastic uncertainty through conditional probability tables (CPTs), yet they lack the capacity to express qualitative assessments such as “likely,” “almost certain,” or “possible.” To bridge this gap, the authors adopt the theory of possibility and necessity, where each proposition is associated with a possibility degree Π (an upper bound) and a necessity degree N (a lower bound) that satisfy Π(A) = 1 − N(¬A). These measures capture a range of belief rather than a single precise probability, offering a more flexible description of incomplete or imprecise information.

The core of the proposed system—named the Fuzzy Possibilistic Bayesian Network (FPBN)—replaces each node’s CPT with a local necessity‑valued knowledge base. For a node i with parent set Pa(i), the knowledge base consists of logical formulas φ_k together with necessity values N(φ_k). These formulas encode expert rules (e.g., “if symptom X is high then disease Y is necessary”) and are interpreted in a fuzzy manner. Crisp certainty degrees supplied by sensors or databases are first fuzzified using membership functions μ(x) that map numeric values to linguistic terms such as “low,” “medium,” or “high.” Consequently, the network operates on fuzzy variables rather than precise probabilities, which improves robustness against noise and missing data.

The inference mechanism proceeds in four stages: (1) fuzzification of input evidence, (2) mapping of fuzzy evidence onto the local necessity knowledge bases, (3) propagation of possibility and necessity values throughout the graph, and (4) application of a decision rule based on the necessity of the target variable exceeding a domain‑specific threshold. Propagation relies on min‑max operations that are computationally cheap; the overall complexity is O(n · k) where n is the number of nodes and k the average number of rules per node, a substantial reduction compared with the exponential blow‑up of exact Bayesian inference.

To keep the model scalable, the authors propose knowledge‑base compression techniques. Redundant or subsuming rules are merged, and necessity intervals are tightened using the minimum and maximum bounds, which reduces both memory consumption and computational load. This compression is analogous to CPT reduction in conventional Bayesian networks but gains additional savings from the fuzzy‑possibilistic dimension.

Empirical evaluation is performed on three real‑world domains: (i) cardiac disease diagnosis, (ii) credit‑risk assessment, and (iii) smart‑manufacturing process control. Each dataset is deliberately corrupted with 10‑30 % missing values and injected Gaussian noise. FPBN is benchmarked against (a) a standard Bayesian network, (b) a fuzzy Bayesian network (FBN) that only fuzzifies inputs, and (c) a pure possibilistic Bayesian network (PBN) that uses possibility/necessity but retains crisp probabilities. Results show that FPBN achieves an average accuracy of 87.3 %, outperforming the conventional Bayesian network (78.5 %) by roughly 9 percentage points. The advantage widens to 12 pp when missing data reach 30 %. Moreover, inference time remains low (≈0.42 s per query), satisfying real‑time requirements, and the system’s performance degrades minimally as noise levels increase, confirming its robustness.

The paper discusses practical implications. In medical settings, clinicians can encode statements like “high fever makes severe infection highly necessary” directly into the network, preserving expert intuition while still benefiting from probabilistic reasoning. In finance, risk analysts can express “the market is possibly volatile” as a possibility interval, enabling more nuanced portfolio decisions. In Industry 4.0, sensor readings that are inherently noisy can be fuzzified, allowing the control system to maintain stability even when data are incomplete.

Nevertheless, the authors acknowledge limitations. Designing appropriate membership functions and initializing necessity values rely heavily on domain expertise, introducing subjectivity. Moreover, as the network grows, the number of fuzzy‑possibilistic rules may increase dramatically, potentially offsetting the gains from compression. Future work is outlined in three directions: (1) automatic learning of membership functions using deep meta‑learning, (2) dynamic adaptation of network structure where nodes and rules are added or pruned in response to streaming data, and (3) integration with multi‑criteria decision analysis to handle trade‑offs among cost, time, and risk.

In summary, this research presents a comprehensive and theoretically grounded approach that fuses Bayesian probabilistic reasoning, fuzzy set theory, and possibilistic logic. By doing so, it delivers a knowledge‑representation framework capable of handling both stochastic uncertainty and linguistic vagueness, offering improved accuracy, robustness, and interpretability for decision‑making systems operating under imperfect information.