TL;DR
This paper introduces a novel clustering-based approach to improve the Dempster-Shafer theory (DST) for combining highly conflicting pieces of evidence. By first clustering similar evidence sources and then combining them, the method aims to overcome DST's counterintuitive results when faced with strong disagreements, leading to more robust and accurate decision-making.
What Problem Does This Paper Address?
The Dempster-Shafer theory (DST) is a powerful framework for reasoning under uncertainty, widely used in fields like multisensor fusion and fault diagnosis. However, a significant challenge arises when DST combines highly conflicting pieces of evidence. In such scenarios, DST can produce counterintuitive results, where the combined evidence strongly supports a conclusion that none of the individual pieces of evidence supported. This "one ticket veto" phenomenon, as illustrated in Example 1, can lead to flawed decisions. The authors aim to address this by developing a method that can effectively handle and mitigate the negative impacts of conflicting evidence before applying DST.
Methods
The proposed method involves several key steps:
- Novel Similarity Measure: A new similarity measure is introduced that considers not only the focal elements and their basic belief assignments (BBAs) but also the order of these assignments. This addresses a limitation in existing measures that might miss "self-conflict" within an evidence source. This measure is further enhanced by incorporating the Pignistic probability distance.
- Evidence Clustering (ISODATA): The ISODATA clustering algorithm is adapted to group similar evidence sources. Instead of traditional Euclidean distance, the custom similarity measure is used to define distances between evidence sources. This groups similar opinions together, reducing the conflict within each cluster.
- Evidence Amendment: Before clustering, each evidence source is amended based on its "commonality" with other sources. This commonality is calculated using the new similarity measure and aims to adjust the evidence based on its support from others.
- Intra-Cluster Combination: Within each cluster, where evidence is assumed to be less conflicting, the Dempster-Shafer rule is applied to combine the evidences and obtain an "evidential representative" for that cluster.
- Inter-Cluster Combination: The evidential representatives from different clusters, which are likely to be more conflicting, are then combined using Murphy's method. This method amends the evidential models without changing the core combination rule, providing a more stable way to combine potentially disparate cluster representatives.
Datasets
The paper does not introduce new datasets. Instead, it uses numerical examples (Examples 3, 4, 5, and 6) to demonstrate the effectiveness of the proposed method. These examples involve synthetic evidence sources with varying degrees of conflict and neutrality to simulate real-world scenarios.
Key Insights and Ideas
- Conflict is not just between sources, but within: The paper highlights that conflict can arise not only between different evidence sources but also from the internal structure of a single evidence source (e.g., the ordering of belief assignments).
- Clustering as a conflict mitigation strategy: By grouping similar evidence first, the method effectively reduces the overall conflict that the Dempster-Shafer rule has to handle, thereby mitigating its counterintuitive behavior.
- A multi-stage approach is superior: The combination of a novel similarity measure, evidence amendment, clustering, and then a two-stage combination (intra-cluster and inter-cluster) provides a more robust solution than direct application of DST to all evidences.
- Reliability matters: The method incorporates a measure of clustering reliability, which considers both the commonality of evidence within a cluster and the number of evidences in that cluster, to weight the contributions of different clusters.
Key Results
The paper presents several numerical examples demonstrating the superiority of their proposed method over existing approaches like the standard Dempster-Shafer (D-S) rule, Yager's method, and others.
- Example 4 (One Conflictive Evidence): The proposed method achieved a significantly higher belief in the true proposition (m(A) = 0.9896) compared to other methods, which often resulted in near-zero belief due to the conflicting evidence. The paper claims high accuracy, strong anti-interference ability, and fast convergence speed.
- Example 5 (One Conflictive and Neutral Evidence): The method showed a more reasonable and intuitive increase in belief for the true proposition when a neutral evidence was introduced, unlike some methods that exhibited decreasing belief.
- Example 6 (Two Conflictive Evidences): The proposed method produced results that were more aligned with intuition, particularly in scenarios where multiple pieces of evidence supported a particular outcome.
Limitations and Future Work
- Parameter Sensitivity: The authors acknowledge that the final combinational result might differ based on the parameters set for the clustering algorithm, especially when the clustering feature is fuzzy. They suggest that a "better clustering approach is expected to reduce this impact in the future."
- Further Improvements: The paper suggests that "more factors can be considered to calculate the reliability of the evidences," implying that the current reliability metric could be further refined.
- Generalizability: While the method is demonstrated with numerical examples, further validation on real-world, complex datasets would be beneficial.
Summary updated