Belief Function Theory and Probabilistic Inference

Belief Function Theory provides a versatile framework for representing and combining uncertain evidence beyond the confines of classical probability. It assigns support to sets of hypotheses via mass functions, accommodating both precise and imprecise information and explicitly quantifying conflict and ignorance. Probabilistic inference encompasses methods for updating beliefs in light of new data, ranging from Bayesian updating and Monte Carlo sampling to variational approximations. The synthesis of belief functions with probabilistic inference has spawned hybrid approaches that integrate likelihood-based updates, entropy-driven constructs and data-driven discounting. These developments enable robust decision making in domains such as sensor fusion, machine learning, risk assessment and legal reasoning. By treating aleatory and epistemic uncertainty separately, belief function theory addresses scenarios where prior knowledge is partial or conflicting. Recent efforts concentrate on scalable combination rules for dependent sources, principled handling of low-quality or incomplete data, and embedding belief assignments within broader statistical pipelines, thus enhancing resilience to model misspecification and measurement errors.

Research from Nature Portfolio

No recent Nature Portfolio content available.

Research from all publishers

Advances in the combination of dependent and partially reliable evidence have been achieved by extending belief-function operators for Gaussian random fuzzy numbers. By introducing correlation-aware combination rules and reliability-sensitive discounting, these methods enhance predictive accuracy in machine learning applications when input sources vary in trustworthiness. A likelihood-based construction of belief functions has been revisited and axiomatized, demonstrating coherence with the likelihood principle, compatibility with Bayesian updating and the minimal commitment principle. This framework has been further extended to accommodate observations of varying quality through auxiliary variables and contour functions, thus broadening applicability to imprecise or partial datasets. An entropy-based belief function paradigm has emerged for forecasting problems where likelihood functions are intractable or unknown. By deriving mass assignments directly from entropy measures, this approach yields prediction intervals with reliable coverage and offers an alternative when statistical models are highly complex or underspecified.

Belief Function Theory and Probabilistic Inference publication trend

The graph below shows the total number of articles in belief function theory and probabilistic inference across all publications each year (not limited to Nature Index journals).

Technical terms

Belief function: A mapping from subsets of a hypothesis space to support values, quantifying the strength of evidence committed exactly to those subsets.

Basic probability assignment (mass function): A non-negative function over all subsets of the frame of discernment that sums to one, representing direct support for each subset.

Dempster’s rule of combination: A normalised conjunctive operator for fusing independent pieces of evidence by combining intersecting mass values.

Plausibility function: A measure of how well the evidence fails to refute a hypothesis, defined as one minus the belief in its complement.

Discounting: A technique for adjusting mass functions to reflect source reliability, reducing support from less trustworthy evidence.

References

1. Combination of dependent and partially reliable Gaussian random fuzzy numbers. Information Sciences (2024).
2. Likelihood-based belief function: Justification and some extensions to low-quality data. International Journal of Approximate Reasoning (2014).
3. Forecasting Using Information and Entropy Based on Belief Functions. Complexity (2020).