Discounted Belief Fusion (DBF)

• Discounted Belief Fusion (DBF) is a framework for combining evidence from multiple sources by discounting unreliable or conflicting inputs to increase overall uncertainty.
• It employs methods from Dempster–Shafer theory and subjective logic, using metrics like conflict- and divergence-based discounting to dynamically weight each source.
• DBF is applied in multimodal learning, medical imaging, and multi-sensor fault diagnosis, delivering improved conflict detection and calibration in complex environments.

Discounted Belief Fusion (DBF) refers to a family of principled frameworks for multi-source or multimodal evidence combination in the presence of source unreliability and conflict, in which each source’s belief or evidence is discounted by a reliability (discount) factor prior to fusion. DBF strategies operate in the formalism of Dempster–Shafer theory or subjective logic, reallocating belief mass associated with unreliable or conflicting sources into uncertainty, thereby enabling robust uncertainty quantification and improved conflict detection. Key variants include conflict-driven discounting for general fusion (Bezirganyan et al., 2024), contextual discounting in deep neural architectures for medical imaging (Huang et al., 2022), dynamic discounting in computer vision tasks (Cao et al., 2016), and divergence-based DBF for multi-sensor environments (Xiao, 2018).

1. Mathematical Formalism and Rationale

Let $\Theta=\{\theta_1,\ldots,\theta_K\}$ denote the frame of discernment. Each modality, sensor, or classifier provides a mass function $m_i: 2^\Theta\rightarrow[0,1]$ with $\sum_{A\subseteq\Theta} m_i(A) = 1$. In subjective logic, this is represented as an opinion $\omega^i = (b^i,u^i,a)$, where $b^i$ is singleton belief, $u^i$ is uncertainty, and $a$ is the base rate vector, with $\sum_k b^i_k + u^i = 1$ (Bezirganyan et al., 2024).

Discounting incorporates a modality- or context-dependent reliability factor $\eta_i\in[0,1]$, modifying each $m_i$ to

$$m_i^{\eta}(A) = \eta_i \cdot m_i(A),\quad A\subset \Theta, \quad m_i^{\eta}(\Theta) = (1-\eta_i) + \eta_i m_i(\Theta).$$

This adjustment reallocates unreliable mass into the total uncertainty. In a class-dependent context, discount vectors $\beta^h\in [0,1]^K$ satisfy $$^{\beta^h}pl^{(h)}(\omega_k)=(1-\beta^h_k)+\beta^h_k pl^{(h)}(\omega_k)$$ for contour functions (Huang et al., 2022).

The rationale is to prevent unreliable or highly conflicting sources from dominating the fused output, by increasing global epistemic uncertainty ($u^\oplus$) and reducing potentially overconfident spurious beliefs.

2. Conflict and Reliability Quantification

Discount factors are determined either per modality, per class, per instance, or dynamically from the data. Several principal techniques are recognized:

• Conflict-based discounting (Bezirganyan et al., 2024): Compute the degree of conflict $DC(\omega^i, \omega^j)$ between all modality pairs using projected distance ($PD$) and conjunctive certainty ($CC$). Then compute the agreement matrix $A_{ij}$, and for each modality,

$\eta_i = \prod_{j=1}^{V} A_{ij}.$

Lower agreement with others lowers $\eta_i$.

• Class-conditional/contextual discounting (Huang et al., 2022): For label set $\Omega=\{\omega_1,\ldots,\omega_K\}$, assign learned vectors $\beta^h=(\beta^h_1,\ldots,\beta^h_K)$ to each modality $h$, reflecting estimated reliability for each class. These are optimized jointly with model parameters via a differentiable discounted Dice loss.
• PR curve-based reliability (Cao et al., 2016): In dynamic fusion for object detection and image classification, reliability factors are implicitly encoded in $m(T)=\mathrm{prec}(s)$ and $m(NT)=\mathrm{rec}^n(s)$, while $m(T\,\mathrm{or}\, NT)$ is interpreted as the discount to ignorance for uncertain predictions.
• Divergence-based support (Xiao, 2018): Discount factors are derived from support degrees computed via a generalised belief Jensen–Shannon divergence (GBJS). The support of each BBA $m_j$ is thereby based on its divergence from all other sources, with discounting proportional to normalized support.

These mechanisms ensure that discount rates reflect both empirical reliability and structural agreement and are order-invariant where required.

3. Fusion Rules, Order-Invariance, and Algorithmic Steps

Once discounted, the revised masses $m_i^\eta$ or opinions $\omega^{i'}$ are fused using designated rules:

• Generalized Averaging Fusion (Bezirganyan et al., 2024):

$$b^{\oplus} = \frac{\sum_{v=1}^{V} b^{v'} U_{-v}}{\sum_{v=1}^{V} U_{-v}},\quad u^{\oplus} = \frac{V \cdot \prod_{v=1}^{V} u^{v'}}{\sum_{v=1}^{V} U_{-v}},\quad\text{where } U_{-v} = \prod_{i\neq v} u^{i'}.$$

This approach is order-invariant, commutative, and associative, and it increases uncertainty mass under high conflict.

• Dempster’s Combination after discounting (Huang et al., 2022, Cao et al., 2016, Xiao, 2018): After discounting, combine masses using

$$m'(A) = \frac{1}{1-\kappa}\sum_{B \cap C = A} {m_1'(B)m_2'(C)},\quad \kappa = \sum_{B\cap C = \emptyset} m_1'(B)m_2'(C).$$

For $H$ modalities, the combination generalizes via sequential or parallel convolutions.

• Weighted-Average Evidence fusion (Xiao, 2018):

$WAE(A) = \sum_{j=1}^k \widetilde{Sup}_j m_j(A),$

where $\widetilde{Sup}_j$ is the normalized support derived from GBJS divergence.

Algorithmic procedures involve computing conflict matrices, agreement factors, discount rates, updating belief masses, and applying the selected fusion rule. Time complexity depends on the number of modalities and classes, typically $O(V^2K)$ per sample for the conflict-driven version (Bezirganyan et al., 2024).

4. Theoretical and Empirical Properties

Key properties of modern DBF methods include:

• Conflict-to-Uncertainty Reallocation: Unlike Dempster’s rule, which normalizes away conflict possibly yielding overconfident assignment to spurious hypotheses, DBF reallocates the conflicting mass into uncertainty, especially for high-conflict or unreliable modalities. As shown in Zadeh’s canonical example, DBF produces high $u^\oplus$ when sources sharply disagree (Bezirganyan et al., 2024).
• Order Invariance: Fusion methods built on commutative, associative operations (generalized averaging, weighted averaging) ensure the fused belief is independent of input order.
• Scalability: These approaches remain robust for an arbitrary number of modalities or sensors, overcoming the order sensitivity and associativity failures of classical pairwise averaging (Bezirganyan et al., 2024).
• Superior Conflict Detection: In multimodal experiments (e.g., HandWritten, CUB, Caltech101, PIE, Scene15), DBF achieves AUC values in conflict-detection tasks substantially higher than classical and generalized averaging fusion rules. For example, on Caltech101, DBF achieves AUC=1.00, compared to 0.72 (BCF) and 0.55 (GBAF) (Bezirganyan et al., 2024).
• Improved Calibration and Reliability: In deep evidential segmentation, learned class-wise discount rates improve Dice, Hausdorff, and ECE metrics (e.g., nnUNet+DBF on BraTS21: Dice ≈90.1%) while yielding per-class reliability vectors in accord with radiological priors (Huang et al., 2022).

5. Principal Applications

Discounted Belief Fusion is widely applicable in multimodal and multi-sensor contexts where source reliability varies, or modalities may carry conflicting information.

• Multimodal Learning: DBF is used for order-invariant, uncertainty-robust sensor fusion in classification tasks across diverse domains including vision, text, and non-visual modalities (Bezirganyan et al., 2024).
• Medical Image Segmentation: In multi-MR segmentation (BraTS 2021), contextual DBF provides state-of-the-art performance, directly learning class-dependent reliability factors by backpropagation (Huang et al., 2022).
• Object Detection and Classification Fusion: DBF fuses detector window scores with image-level classification priors to reduce false positives and improve mAP in benchmark (VOC07/12) detection tasks (Cao et al., 2016). Here, discounting operates dynamically via surplused “ignorance” mass, with empirical mAP gains up to +0.075 for weak detectors.
• Multi-sensor Fault Diagnosis: Reliability-weighted GBJS discounting enables robust diagnosis in fault-prone industrial settings, adapting the contribution of each sensor based on sufficiency and importance indices (Xiao, 2018).

6. Practical Implementation and Guidelines

• Discount Rate Sensitivity: Hyperparameters (e.g., $\lambda$ in conflict-to-agreement mapping (Bezirganyan et al., 2024)) control how aggressively unreliable modalities are discounted; selection via validation is recommended.
• Stability: In the presence of many modalities, products of agreement factors may underflow; a log-domain implementation or normalization thresholding prevents numerical issues.
• Architecture Integration: End-to-end frameworks can learn discount rates as contextual or class-wise parameters, directly optimizing data-driven objectives (e.g., discounted Dice loss) in neural architectures (Huang et al., 2022).
• Computational Cost: While the order-invariant averaging fusion employed by DBF is scalable, the initial computation of all conflict or support degrees is quadratic in modality count. Efficient parallelization or early pruning of low-contribution modalities can substantially reduce overhead (Bezirganyan et al., 2024).
• Interpretability: Final discount rates in contextual DBF reflect empirical utility: high rates for informative modalities/class pairs and low otherwise, which provides actionable reliability diagnostics and mitigates over-reliance on spurious modalities (Huang et al., 2022).

7. Limitations and Open Directions

• Assumed Uniform Base Rates: Most DBF protocols currently assume uniform priors; extensions to class-dependent or hierarchical priors require new formulations (Bezirganyan et al., 2024).
• Hyperparameter Selection: The mapping sensitivity from conflict to agreement (controlled by $\lambda$) requires tuning; adaptive or data-driven schedules are open challenges.
• Continuous and Hierarchical Frames: Extensions of DBF to non-discrete frames (continuous hypotheses, hierarchies) remain under-explored.
• Class-vs-Instance Discounting: Further research is warranted on dynamic instance-level discounting as opposed to global or class-conditional forms.
• Numerical Stability and Underflow: Product-based reliability aggregations may suffer from underflow as the number of modalities increases, necessitating robust implementation strategies.
• Broader Utility: While substantial performance improvements are demonstrated in multimodal learning, medical imaging, and object detection, direct comparisons with non-evidence-theoretic uncertainty quantification remain limited and represent a direction for further empirical study.

---

Key references: "Multimodal Learning with Uncertainty Quantification based on Discounted Belief Fusion" (Bezirganyan et al., 2024), "Evidence fusion with contextual discounting for multi-modality medical image segmentation" (Huang et al., 2022), "Enhanced Object Detection via Fusion With Prior Beliefs from Image Classification" (Cao et al., 2016), "Multi-sensor data fusion based on a generalised belief divergence measure" (Xiao, 2018).