Trust Transitivity and Conditional Belief Reasoning

Audun Jøsang, Tanja Ažderska, Stephen MarshIFIP 2012 · DOI · OpenAlex

Abstract

This paper presents a formal mathematical framework for trust transitivity by modeling it as a problem of conditional belief reasoning within Subjective Logic. By moving beyond traditional ad hoc mathematical operators, the authors provide a robust, evidence-based approach that accounts for uncertainty and base rates. The proposed model demonstrates that existing trust transitivity operators are restricted special cases of a more general conditional logic framework, offering a more nuanced representation of human trust in computational systems.

Introduction

In the domain of computational trust, transitivity refers to the process by which trust is inferred through a chain of recommendations: if Alice trusts Bob, and Bob recommends Claire, Alice may derive a level of trust in Claire. Traditional models often rely on simplistic operators, such as the product rule, which can lead to erroneous conclusions. Specifically, in long recommendation chains, trust values often converge to zero, which many systems incorrectly interpret as "distrust" rather than "no information." Furthermore, these models frequently fail to distinguish between a recommender's integrity and their expertise. This paper addresses these limitations by applying Subjective Logic to formalize trust transitivity as conditional reasoning.

Methodology

The authors utilize Subjective Logic (SL), a probabilistic logic that explicitly models uncertainty. An opinion in SL is defined by the quadruple $(b, d, u, a)$, representing Belief, Disbelief, Uncertainty, and the Base Rate, respectively.

Geometric Representation of Opinions

Opinions are visualized as points within a barycentric coordinate system. A binomial opinion (binary choice) is represented within an equilateral triangle, while a multinomial opinion (n-ary choice) is represented within a tetrahedron.

Geometric representation of binomial (left) and multinomial (right) opinions.

Conditional Deduction and Sub-simplex Mapping

The core of the methodology is the reinterpretation of recommendations as conditional logic. A recommendation is treated as an antecedent, and the target's trustworthiness as a consequent. This is operationalized through "sub-simplex mapping," where a conditional opinion defines a sub-region (sub-simplex) within the target frame.

Conditional mapping for binomial opinions, illustrating the projection from a parent frame to a child frame.

The framework also differentiates between causal and derivative conditionals, using Bayesian inversion to avoid common pitfalls like the base rate fallacy.

Key Insights

  • Evidence-Based Transitivity: Trust transitivity is framed as an analysis of evidence rather than a propagation of trust. This strips away the "mysticism" of transitivity, treating it as a standard logical deduction.
  • Uncertainty Modeling: By incorporating an uncertainty parameter ($u$), the model distinguishes between a lack of evidence and active disbelief, preventing the "zero trust" error common in product-based models.
  • Unification of Operators: The authors demonstrate that several existing trust operators—including Uncertainty Favoring and Base Rate Sensitive operators—are specific instances of their generalized conditional model.
  • Separation of Concerns: The model allows for the independent assessment of a recommender's honesty and their competence, enabling more complex and realistic trust networks.

Results and Conclusion

The paper provides mathematical proof that the conditional model can replicate the behavior of previous operators by appropriately setting conditional opinions. This flexibility allows the model to handle "negative" transitivity, such as discounting or inverting recommendations from known biased sources.

While powerful, the authors note that the model assumes a high degree of rationality, which may not fully capture the affective or emotional components of human trust. Future research is encouraged to integrate contextual factors, such as situational urgency or physical location, to further enhance the model's applicability to real-world scenarios.

Summary updated

BibTeX
@inproceedings{jsang2012trust,
  author = {Audun Jøsang and Tanja Ažderska and Stephen Marsh},
  title = {{Trust Transitivity and Conditional Belief Reasoning}},
  booktitle = {IFIP},
  year = {2012},
  doi = {10.1007/978-3-642-29852-3\_5},
}
Figure