Conflict-Aware Dempster–Shafer Rules (DSCR)

• Conflict-aware Dempster–Shafer combination rules (DSCR) are evidence fusion methods that explicitly treat conflict instead of normalizing it away.
• DSCR encompasses techniques like conflict redistribution, reweighting, and geometric adjustments to handle sensitivity in classical Dempster–Shafer fusion.
• Applications include multi-sensor fault diagnosis and incomplete multi-view classification, where DSCR improves robustness against counterintuitive fusion under high conflict.

Conflict-aware Dempster–Shafer Combination Rule (DSCR) denotes a class of Dempster-style evidence-fusion procedures in which evidential conflict is treated explicitly rather than being normalized away without qualification. In the material considered here, the term appears in two senses. In a narrow sense, it names the conflict-aware mechanism used in the Alternating Progressive Learning Network (APLN) for incomplete multi-view classification, where subjective-logic opinion fusion is regulated by a Jensen–Shannon-divergence-based conflict loss (Chen et al., 2024). In a broader sense, the literature motivates conflict-aware combination whenever standard Dempster fusion is sensitive to incompatible evidence, undefined under total conflict, dependent on frame refinement and source independence, or prone to counterintuitive decisions under high disagreement (Yen, 2013).

1. Classical baseline and the source of the conflict problem

The baseline object for DSCR is the normalized Dempster combination rule. For two mass functions $m_1$ and $m_2$, the combined mass is

$m(A)=\frac{\sum_{B\cap C = A} m_1(B)m_2(C)}{1-K},$

for nonempty $A$, where the conflict coefficient is

$K = \sum_{B\cap C = \varnothing} m_1(B)m_2(C).$

When the two mass distributions are totally conflicting, the normalization factor is zero, and the rule does not combine them (Yen, 2013).

The central concern is not merely that $K$ can be nonzero, but that normalization can produce anomalous behavior when conflict is high. In multisensor fault diagnosis, an example with two highly conflicting sensors illustrates the issue: sensor 1 strongly supports $F_1$, sensor 2 strongly supports $F_3$, and both weakly support $F_2$, yet classical DCR fuses them into $m(F_1)=0$, $m_2$0, $m_2$1 (Ghosh et al., 2020). In that setting, the normalization denominator $m_2$2 becomes very small, and tiny shared masses can dominate the result.

The literature therefore distinguishes between partial conflict, which may still admit a meaningful combination, and total conflict, where combination may be inadmissible. One formulation states that if there are no non-empty intersections between focal elements, the evidential sources are mutually exclusive and “should not be combined at all” (Yen, 2013). This conflict-sensitive stance is the conceptual starting point for DSCR.

Approach	Conflict treatment	Noted property
Conditional granular distributions (Yen, 2013)	Combine only when Dempster’s rule is applicable	Conflict tied to admissibility
Refined-frame interpretation (1304.1143)	Recast evidence on a refined frame	Preserves independence requirements
PCR6 (0806.1797)	Local proportional conflict redistribution	Non-associative
DCR1/DCR2 for D numbers (Deng et al., 2017)	Use non-exclusive degree and redistribute to intersection/union	Reduce to Dempster’s rule under DST conditions
iDCR (Ghosh et al., 2020)	Reweight BPAs before DCR	Robust under high-conflict sensors
TP/mTP (1304.1125)	Conflict widens intervals geometrically	Associative and commutative
APLN DSCR (Chen et al., 2024)	JS-based conflict loss on opinions	Conflict-aware training in EDL

2. Combinability, admissibility, and evidential interpretation

A recurrent theme in conflict-aware work is that conflict is not only a numerical artifact but also a question of whether combination is semantically justified. In a relational interpretation of Dempster–Shafer theory, Zadeh’s noncombinability concern is addressed by distinguishing unconditioned granular distributions from conditional granular distributions. If masses are treated as unconditioned summaries of a relation, combinability can fail in the presence of conflict-free-parent-relation requirements. If masses are treated as conditional granular distributions, however, the masses are combinable whenever Dempster’s rule is applicable (Yen, 2013).

That distinction is formalized through a theorem. For conditional granular distributions

$m_2$3

and

$m_2$4

if their focal elements have at least one non-empty intersection, then there exists at least one conflict-free parent relation for the combined conditional granular distribution $m_2$5 (Yen, 2013). Conflict-awareness here is therefore a condition on the interpretation of evidence sources and the existence of a compatible parent relation.

A second line of argument locates the problem in frame construction and independence assumptions rather than in Dempster’s rule alone. In non-monotonic reasoning, evidence such as “birds normally fly,” “penguins are birds,” and “penguins normally do not fly” becomes counterintuitive under naive fusion because the rules are not being represented on the same refined frame and are not independent in the naive Dempster sense (1304.1143). The proposed remedy is to refine the frame from $m_2$6 to

$m_2$7

where $m_2$8 denotes flying penguins, $m_2$9 non-flying penguins, $m(A)=\frac{\sum_{B\cap C = A} m_1(B)m_2(C)}{1-K},$0 other flying birds, and $m(A)=\frac{\sum_{B\cap C = A} m_1(B)m_2(C)}{1-K},$1 other non-flying birds (1304.1143).

On that refined frame, the bird rule is represented as

$m(A)=\frac{\sum_{B\cap C = A} m_1(B)m_2(C)}{1-K},$2

and penguin-negative evidence as

$m(A)=\frac{\sum_{B\cap C = A} m_1(B)m_2(C)}{1-K},$3

The resulting belief interval for $m(A)=\frac{\sum_{B\cap C = A} m_1(B)m_2(C)}{1-K},$4 is

$m(A)=\frac{\sum_{B\cap C = A} m_1(B)m_2(C)}{1-K},$5

so the belief that penguins fly remains small, while the belief that other birds fly remains strong (1304.1143). In this formulation, conflict-awareness is achieved by modeling discipline: refine the frame, preserve source independence, and only then combine.

3. Conflict redistribution and nonexclusive generalizations

One major DSCR strategy is explicit redistribution of conflicting mass. In the proportional conflict redistribution literature, PCR6 is defined by first computing the conjunctive consensus, identifying each conflicting partial product, and redistributing that conflicting mass only among the hypotheses directly involved in the conflict, proportionally to their own masses (0806.1797). The rule is intended to be more local than Dempster’s global normalization and more decisive than the Dubois–Prade rule in many cases. Its stated drawback is non-associativity, which can be problematic in dynamic or sequential fusion (0806.1797).

The same conflict-redistribution perspective appears in D numbers. The abstract of D numbers theory states that the approach was proposed to overcome shortcomings in Dempster–Shafer theory caused by the exclusiveness hypothesis and the completeness constraint, and that it generalizes Dempster–Shafer theory while strengthening uncertainty modelling (Deng et al., 2014). A later paper develops two explicit combination rules for D numbers, DCR1 and DCR2, from a conflict-redistribution viewpoint (Deng et al., 2017).

A D number is a mapping

$m(A)=\frac{\sum_{B\cap C = A} m_1(B)m_2(C)}{1-K},$6

such that

$m(A)=\frac{\sum_{B\cap C = A} m_1(B)m_2(C)}{1-K},$7

The quantity

$m(A)=\frac{\sum_{B\cap C = A} m_1(B)m_2(C)}{1-K},$8

is the degree of completeness, and unlike DST the elements of $m(A)=\frac{\sum_{B\cap C = A} m_1(B)m_2(C)}{1-K},$9 are not required to be mutually exclusive (Deng et al., 2017).

The conflict-aware ingredient is the non-exclusive degree

$A$0

defined by

$A$1

with symmetry $A$2 (Deng et al., 2017).

For complete D numbers, DCR1 is

$A$3

where

$A$4

For incomplete D numbers, DCR2 is

$A$5

with

$A$6

The factor $A$7 must satisfy

$A$8

These rules reduce to classical Dempster’s rule when completeness holds and the non-exclusiveness of disjoint propositions vanishes (Deng et al., 2017).

This suggests a broad DSCR principle: disjointness need not always be treated as absolute contradiction. In the D-number setting, a disjoint pair can contribute partly to residual conflict and partly to the union $A$9, depending on $K = \sum_{B\cap C = \varnothing} m_1(B)m_2(C).$0 (Deng et al., 2017).

4. Reweighting and geometric rules for robust fusion

Another DSCR strategy is to alter the inputs to Dempster fusion rather than the normalization alone. In iDCR for multisensor fault diagnosis, each sensor’s BPA is reweighted before fusion using two quantities: how well that sensor agrees with the other sensors, and how uncertain that sensor’s evidence is, measured by weighted Deng entropy (Ghosh et al., 2020).

Given evidence set $K = \sum_{B\cap C = \varnothing} m_1(B)m_2(C).$1, the algorithm computes the average BPA

$K = \sum_{B\cap C = \varnothing} m_1(B)m_2(C).$2

the Euclidean distance

$K = \sum_{B\cap C = \varnothing} m_1(B)m_2(C).$3

the similarity

$K = \sum_{B\cap C = \varnothing} m_1(B)m_2(C).$4

and the support degree

$K = \sum_{B\cap C = \varnothing} m_1(B)m_2(C).$5

Uncertainty is measured by weighted Deng entropy,

$K = \sum_{B\cap C = \varnothing} m_1(B)m_2(C).$6

and credibility weights are then defined by

$K = \sum_{B\cap C = \varnothing} m_1(B)m_2(C).$7

The modified BPA is

$K = \sum_{B\cap C = \varnothing} m_1(B)m_2(C).$8

after which classical DCR is applied to the weighted evidence (Ghosh et al., 2020).

The numerical effect is explicitly conflict-aware. For the highly conflicting example

$K = \sum_{B\cap C = \varnothing} m_1(B)m_2(C).$9

classical DCR yields $K$0, $K$1, $K$2, whereas iDCR yields

$K$3

(Ghosh et al., 2020). Here, conflict is moderated through reliability-aware discounting.

A geometrically different approach appears in interval-based evidence combination. Evidence is represented by an interval $K$4, and the “discrimination” quantity

$K$5

determines whether the interval leans toward $K$6 or toward $K$7 (1304.1125). Two intervals are conflicting if their discrimination values have opposite signs. The Triangle-to-Plane (TP) rule maps the triangular evidence region into a plane, combines transformed coordinates by simple addition,

$K$8

and maps back. The method is designed to satisfy closure, commutativity, associativity, continuity, identity, and symmetry, while also ensuring that conflict increases uncertainty and agreement decreases uncertainty (1304.1125).

The width of the resulting interval is

$K$9

so opposing $F_1$0-components reduce the vector norm and widen the interval under conflict (1304.1125). In the paper’s example, inputs $F_1$1 and $F_1$2 are intuitively in conflict; D-S yields $F_1$3, while TP yields $F_1$4 (1304.1125). The modified TP rule, mTP, extends this construction to dependent evidence by replacing simple addition with a parameterized combining function $F_1$5 (1304.1125).

5. The DSCR mechanism in uncertainty-aware incomplete multi-view classification

In APLN, the phrase “conflict-aware Dempster-Shafer combination rule” refers to a specific mechanism embedded in an Evidential Deep Learning pipeline for incomplete multi-view classification (Chen et al., 2024). The paper states that vanilla Dempster-Shafer-style fusion is highly sensitive to conflicting evidence and that even a single piece of conflicting evidence can lead to anomalous fusion results. The proposed DSCR is not a completely new closed-form fusion operator that replaces the standard rule in all places; rather, it introduces a conflict-aware consistency loss built on a conflict degree measure derived from Jensen–Shannon divergence, together with a subjective-logic-style opinion aggregation rule (Chen et al., 2024).

Each view produces a subjective logic opinion

$F_1$6

where $F_1$7 is belief mass over classes, $F_1$8 is uncertainty mass, and $F_1$9 is the prior, with

$F_3$0

Projected class probabilities are

$F_3$1

For two views $F_3$2 and $F_3$3, the paper defines the aggregated opinion

$F_3$4

with

$F_3$5

$F_3$6

and

$F_3$7

Multi-view fusion is written as

$F_3$8

These are the combination formulas used for dependent opinions in the method (Chen et al., 2024).

The conflict-aware component is defined on uncertainty-adjusted distributions

$F_3$9

followed by Jensen–Shannon divergence

$F_2$0

with

$F_2$1

The conflict degree is

$F_2$2

and the conflict loss is

$F_2$3

The evidential objective is

$F_2$4

and joint training uses

$F_2$5

The paper states that the conflict degree measure is symmetric, stays in $F_2$6, uses $F_2$7 for maximum conflict and $F_2$8 for no conflict, and unifies probability and uncertainty into one metric (Chen et al., 2024).

The DSCR mechanism is integrated into three stages: UMAE-F, UMAE-V, and UMAE-J. Conflict awareness enters most directly in UMAE-V and continues to shape latent representations during joint training (Chen et al., 2024). On conflict datasets created by randomly replacing one view in 40% of samples with a view from another category, the paper reports that the conflict measurement method effectively captures the conflicts, that the conflict loss significantly enhances robustness, and that it effectively reduces conflict risks in the DSCR (Chen et al., 2024).

6. Critiques, misconceptions, and unresolved technical issues

A persistent misconception is that conflict-aware DSCR merely means “another normalization formula.” Several papers argue that the underlying issue is broader: the meaning of masses, the admissibility of combination, and the relation between belief functions and probability all become problematic under conflict (Brodzik et al., 2011). In one two-source example,

$F_2$9

with masses

$m(F_1)=0$0

Dempster’s rule yields

$m(F_1)=0$1

whereas the corresponding probabilities from the law of total probability are

$m(F_1)=0$2

The paper concludes that, in the general case with non-zero conflict $m(F_1)=0$3, DS-combined masses need not agree with the corresponding probabilities (Brodzik et al., 2011).

A stronger critique argues that the conjunction of four assertions is inconsistent: chance is the limit of relative frequency; Bayesian belief function equals chance; weights of evidence from distinct sources add; and Dempster’s rule is the correct evidence-combination rule (Wang, 2013). In the binary frame $m(F_1)=0$4, repeated combination under Dempster’s rule drives the limiting belief to

$m(F_1)=0$5

rather than to the actual chance $m(F_1)=0$6 (Wang, 2013). This is offered as a defect in the theory’s treatment of probability-like beliefs.

Conflict-aware proposals also carry their own limitations. PCR5 and PCR6 are non-associative (0806.1797). The D-number rules require prior specification of the non-exclusive degree matrix $m(F_1)=0$7 and the function $m(F_1)=0$8, and the combination is not associative (Deng et al., 2017). The APLN paper does not provide a formal proof of correctness or convergence for DSCR itself; its theoretical discussion is described as fairly high-level (Chen et al., 2024). iDCR is presented as effective empirically, but future work is said to be needed for more dynamic and complex environments (Ghosh et al., 2020).

Taken together, these works suggest that conflict-aware DSCR is best understood not as a single universally accepted operator but as a technical design space. Within that space, conflict may be treated as a signal of inadmissible combination, as a quantity to be redistributed locally, as a symptom of nonexclusive hypotheses or incomplete information, as a reliability cue for reweighting, or as a training-time regularizer in learned evidential fusion (Yen, 2013).