Dempster-Shafer Fusion
Dempster-Shafer Evidence Fusion · Also known as: belief function fusion, evidence combination
Dempster-Shafer fusion is an ensemble method based on evidence theory (belief functions) that combines predictions from multiple sources by assigning basic probability masses to subsets of hypotheses. Rather than requiring a probability distribution over single outcomes, it allows uncertainty over sets of outcomes, providing a richer representation of confidence and doubt. Developed by Dempster (1968) and formalized by Shafer (1976), this method is particularly useful when sources are unreliable, conflicting, or provide partial evidence.
Read the full method
Sign in with a free account to read this section.
Method map
The neighbourhood of related methods — select a node to explore.
When to use it
Use Dempster-Shafer fusion when classifiers are unreliable, provide uncertain predictions, or contain conflicting evidence. It excels in multi-source data fusion (e.g., combining radar, infrared, and visual sensors) and diagnostic tasks where uncertainty must be explicitly represented. Avoid it if classifiers are highly accurate and provide hard predictions, as simpler methods (majority voting) suffice and are computationally cheaper.
Strengths & limitations
- Captures uncertainty over sets of hypotheses, not just single outcomes.
- Handles conflict between sources explicitly through the conflict measure K.
- Theoretically sound framework for combining uncertain evidence from multiple sources.
- Distinguishes between uncertainty (lack of information) and probability (likelihood of outcomes).
- Works well with heterogeneous or unreliable classifiers.
- Computationally expensive: combining N classifiers over M hypotheses requires O(2^M) operations.
- Sensitive to mass assignment: converting classifier outputs to masses requires careful calibration.
- High conflict (K close to 1) indicates fundamentally incompatible sources; the method provides no principled resolution.
- Requires more domain knowledge than simple voting methods; choosing mass functions is non-trivial.
Frequently asked
How do I assign basic probability masses from classifier outputs?
Convert confidence scores to masses: if a classifier predicts class A with confidence p, assign m({A})=p and m({all hypotheses})=1-p. For soft predictions, more sophisticated methods map probability distributions to focal elements (subsets of specific interest).
What does the conflict measure K represent?
K measures the degree of conflict between sources. K=0 means agreement; K>0.5 indicates severe conflict. High conflict suggests sources are incompatible or fundamentally contradictory. If K > 0.7, consider removing conflicting sources or validating their reliability.
Should I use belief, plausibility, or average?
Belief is optimistic (sums only certain evidence); plausibility is pessimistic (sums possible evidence). For conservative decisions, use belief; for optimistic ones, use plausibility. Many applications use the average or find the class with lowest plausibility of disbelief (1-plausibility).
How does Dempster-Shafer differ from Bayesian fusion?
Bayesian methods require a prior distribution and likelihood; Dempster-Shafer avoids priors by working with belief and plausibility. Dempster-Shafer handles ignorance better (mass over {all hypotheses}), while Bayesian fusion is more efficient computationally.
Sources
- Dempster, A. P. (1968). A generalization of Bayesian inference. Journal of the Royal Statistical Society, 30(2), 205-247. DOI: 10.1111/j.2517-6161.1968.tb00722.x ↗
- Shafer, G. (1976). A Mathematical Theory of Evidence. Princeton University Press. link ↗
How to cite this page
ScholarGate. (2026, June 3). Dempster-Shafer Evidence Fusion. ScholarGate. https://scholargate.app/en/ensemble-learning/dempster-shafer-fusion
Which method?
Set this method beside its closest kin and read them side by side — the library lays the books on the table; the choice is yours.
- Majority VotingEnsemble Learning↔ compare
- WEIGHTED-VOTINGDecision-making↔ compare