This paper introduces a unified evidence reasoning framework designed to improve how multi-source information is combined under Dempster-Shafer theory (DST). It addresses two primary weaknesses in existing fusion methods: the inability to jointly evaluate conflict and uncertainty, and the reliance on instantaneous comparisons that ignore the long-term reliability of evidence sources.
Addressing Fusion Challenges
Traditional DST-based fusion often struggles when evidence sources provide conflicting or imprecise information. Existing conflict measures typically treat inter-evidence inconsistency and intra-evidence uncertainty as separate issues, leading to incomplete evaluations. Furthermore, current methods evaluate sources based only on their immediate performance, failing to account for how a source has behaved across different decision contexts in the past. This can lead to the suppression of generally reliable sources that are occasionally conflicting, or the failure to penalize systematically unreliable ones.
The Proposed Framework
The authors propose a framework consisting of two core mechanisms:
Chaos-Conflict Measurement (CCM): This tool provides a single scalar quantity that evaluates both cross-evidence association and intra-evidence non-specificity. By doing so, it captures the internal structure of multi-element focal elements—which represent ignorance—rather than just measuring total disagreement. The authors provide formal proofs for five properties of this measure, including boundedness and symmetry, to ensure consistent assessment.
Historical-Experience Weighting: This scheme uses spectral clustering to partition the decision space into specific contexts. It then applies regret theory to analyze past fusion outcomes. By calculating "regret" (when an evidence source steers an ensemble away from the truth) and "rejoice" (when it pulls the ensemble toward the truth), the framework assigns context-specific reliability weights to each source based on their long-term track record.
Performance and Results
The framework integrates these mechanisms into a hybrid combination rule that adaptively balances uncertainty preservation against weighted consensus, depending on the level of global conflict. A belief-interval decision strategy is then used for classification.
In experiments conducted on 16 real-world benchmark datasets, the framework achieved an average F1 score of 85.78 and a mean AUC of 93.30. These results outperformed eight DST-based baselines and three gradient boosting methods. Ablation analysis indicated that the historical experience weighting component was the most significant contributor to performance, with its removal resulting in a 5.03% reduction in AUC.
Considerations for Implementation
The framework is designed for multi-source decision-making environments where evidence sources are used repeatedly across heterogeneous contexts. While the authors demonstrate that the approach is robust against feature noise, label noise, and varying numbers of evidence sources, the effectiveness of the historical weighting relies on the availability of past fusion outcomes to build reliability profiles. The authors note that this approach is applicable to any setting where evidence sources operate across diverse, recurring decision scenarios.
Comments (0)
to join the discussion
No comments yet
Be the first to share your thoughts!