Biased Evidential Multi-View Learning

• Biased Evidential Multi-view Learning (BEML) is a framework that addresses systematic bias in evidence allocation across different views to improve fairness in classification.
• FAML integrates an adaptive prior, fairness regularization, and cross-view opinion alignment to prevent confident misclassification and high uncertainty in minority classes.
• Empirical results across multiple datasets show that incorporating these strategies boosts accuracy and lowers uncertainty estimation errors compared to baseline evidential methods.

Searching arXiv for the cited papers to ground the article in the current literature. Biased Evidential Multi-view Learning (BEML) denotes a problem setting in multi-view evidential learning in which the core failure arises not only from disagreement during fusion, but from biased evidence allocation within each view. In the formulation introduced by "Fairness-Aware Multi-view Evidential Learning with Adaptive Prior" (Chen et al., 18 Aug 2025), samples from data-rich classes tend to receive more evidential support, while minority or less-supported classes may be assigned evidence toward the wrong classes. The result is a joint degradation of classification behavior and uncertainty quality: minority-class instances can be confidently misclassified, and even correct minority-class predictions can remain highly uncertain. BEML therefore concerns the learning of view-specific evidence, the conversion of that evidence into Dirichlet opinions, and the fusion of those opinions under a fairness requirement that evidence assigned to the ground-truth class should be comparable across classes.

1. Problem setting and empirical motivation

The BEML formulation begins from the conventional multi-view evidential learning assumption that each view learns evidence reliably and that subsequent fusion is the main technical challenge. The motivating paper argues that this assumption is often violated in practice. Its empirical study on the BRCA dataset reports that samples from a minority or less-supported class such as Her2 are often assigned evidence toward majority or data-rich classes such as Normal or Basal. The error is therefore not merely inter-view conflict; it is a systematic distortion in how evidence is generated within individual views (Chen et al., 18 Aug 2025).

This distortion produces two distinct failure modes. First, minority-class samples may receive strong evidence for the wrong class, leading to confident misclassification. Second, correct predictions for minority classes may receive too little evidence, yielding high predictive uncertainty even when the label is correct. In that sense, BEML is not reducible to ordinary class imbalance at the level of logits or posterior probabilities. Its defining object is the evidential allocation process itself.

The formal setup considers a multi-view dataset

$D=\{\{x_n^v\}_{v=1}^V, y_n\}_{n=1}^N,$

where $x_n^v \in \mathbb{R}^{d_v}$ is the feature vector of the $v$-th view and $y_n \in \{1,\dots,K\}$ is the class label. The learning objective is to obtain view-specific evidence, convert that evidence into Dirichlet opinions, and fuse the resulting opinions into a final prediction with uncertainty. The fairness-aware requirement proposed for BEML is

$$\mathbb{E}_{(x, y=k) \in D}[e_k^v(x)] \approx \mathbb{E}_{(x, y=k') \in D}[e_{k'}^v(x)] \qquad \forall k,k' \in \{1,\dots,K\}.$$

This criterion states that, for each view $v$, the expected evidence assigned to the ground-truth class should not systematically favor data-rich classes over data-poor ones.

2. Evidential representation and uncertainty semantics

In the evidential multi-view framework used to define BEML, each view-specific network outputs a nonnegative evidence vector

$\mathbf{e} = \{e_1,\dots,e_K\}.$

These evidence values parameterize a Dirichlet distribution through

$\alpha_k = e_k + 1.$

Using subjective logic, the corresponding belief masses and uncertainty mass are

$$b_k = \frac{e_k}{S} = \frac{\alpha_k - 1}{S}, \qquad u = 1 - \sum_{k=1}^K b_k = \frac{K}{S},$$

with Dirichlet strength

$S=\sum_{k=1}^K \alpha_k.$

The projected class probability is then

$x_n^v \in \mathbb{R}^{d_v}$0

where $x_n^v \in \mathbb{R}^{d_v}$1 is the base rate, usually uniform in standard evidential deep learning (Chen et al., 18 Aug 2025).

Under this representation, larger evidence increases $x_n^v \in \mathbb{R}^{d_v}$2 and therefore decreases $x_n^v \in \mathbb{R}^{d_v}$3. The central difficulty in BEML is that low uncertainty is not necessarily trustworthy when the evidence itself is biased. A minority-class sample may accumulate large evidence in favor of a majority class, and the model then becomes confidently wrong. Conversely, a correct minority-class decision may still carry little evidence and therefore appear unreliable. The BEML diagnosis is thus targeted at the calibration of evidential support, not solely at post hoc confidence estimation.

A common misconception is to treat BEML as a fusion-only pathology. The formulation explicitly rejects that view. The motivating argument is that a multi-view model can fuse opinions correctly under its chosen operator and still fail because the opinions being fused already encode biased class support.

3. Fairness-Aware Multi-view Evidential Learning (FAML)

The principal method proposed for BEML is Fairness-Aware Multi-view Evidential Learning (FAML), which is organized in two stages. The first stage performs view-specific evidential learning, producing evidence vectors and corresponding Dirichlet opinions per view. The second stage fuses these opinions through uncertainty-aware weighted aggregation and then imposes an opinion alignment mechanism so that inconsistent or biased opinions are discouraged during multi-view integration (Chen et al., 18 Aug 2025).

A central component of FAML is a training-trajectory-based adaptive prior, introduced in place of the fixed uniform Dirichlet prior used in standard evidential learning. For class $x_n^v \in \mathbb{R}^{d_v}$4, the adaptive prior is defined as

$x_n^v \in \mathbb{R}^{d_v}$5

where $x_n^v \in \mathbb{R}^{d_v}$6 is the number of samples in class $x_n^v \in \mathbb{R}^{d_v}$7, $x_n^v \in \mathbb{R}^{d_v}$8 is a hyperparameter controlling prior strength, $x_n^v \in \mathbb{R}^{d_v}$9 is the model prediction, and

$v$0

The adjusted Dirichlet concentration becomes

$v$1

The adaptive prior is intended to depend on how well the model has been predicting each class over training. If a class is underperforming, the prior changes accordingly. In the paper’s implementation details, the prior is not used from the start; it is activated after 20 epochs and then updated every 5 epochs. This staging is presented as a way to avoid unstable early predictions from corrupting the prior.

Functionally, the adaptive prior acts as a regularization strategy for evidence learning. The paper characterizes it as a mechanism that calibrates the biased evidence learning process, prevents dominant classes from overwhelming the posterior too early, and helps minority classes receive more balanced support. A plausible implication is that FAML treats the prior not as a static epistemic baseline, but as a dynamic control variable coupled to class-wise learning dynamics.

4. Fairness regularization and cross-view opinion alignment

FAML makes the fairness objective explicit through a fairness degree defined as the variance of class-wise average evidence:

$v$2

where $v$3 is the average evidence for class $v$4 and $v$5 is the global average over classes. Low $v$6 indicates more even evidence distribution across classes; high $v$7 indicates uneven concentration and therefore bias (Chen et al., 18 Aug 2025).

This quantity is incorporated into the classification objective:

$v$8

where the expected cross-entropy under the Dirichlet is

$v$9

The coefficient $y_n \in \{1,\dots,K\}$0 gradually increases from 0 to 1 during training, so the fairness constraint becomes progressively stronger as optimization proceeds.

FAML also addresses bias at the fusion stage through opinion alignment. For a view $y_n \in \{1,\dots,K\}$1, the variance of the Dirichlet parameter for class $y_n \in \{1,\dots,K\}$2 is given as

$y_n \in \{1,\dots,K\}$3

The dissonance between two views $y_n \in \{1,\dots,K\}$4 and $y_n \in \{1,\dots,K\}$5 is

$y_n \in \{1,\dots,K\}$6

and the alignment regularizer is

$y_n \in \{1,\dots,K\}$7

The stated purpose is to reduce view-specific bias, discourage a misleading high-confidence opinion from dominating fusion, and encourage mutually supportive evidence across views.

Before final fusion, FAML uses confidence-weighted evidence aggregation. For two views $y_n \in \{1,\dots,K\}$8 and $y_n \in \{1,\dots,K\}$9,

$$\mathbb{E}_{(x, y=k) \in D}[e_k^v(x)] \approx \mathbb{E}_{(x, y=k') \in D}[e_{k'}^v(x)] \qquad \forall k,k' \in \{1,\dots,K\}.$$0

where $$\mathbb{E}_{(x, y=k) \in D}[e_k^v(x)] \approx \mathbb{E}_{(x, y=k') \in D}[e_{k'}^v(x)] \qquad \forall k,k' \in \{1,\dots,K\}.$$1 and $$\mathbb{E}_{(x, y=k) \in D}[e_k^v(x)] \approx \mathbb{E}_{(x, y=k') \in D}[e_{k'}^v(x)] \qquad \forall k,k' \in \{1,\dots,K\}.$$2 are the view confidences. This extends to all views and yields a joint opinion.

5. Optimization protocol and empirical results

The overall training objective combines evidence classification loss with fairness regularization, view-wise losses across all views, opinion alignment loss, and class-balanced weighting. The paper states that $$\mathbb{E}_{(x, y=k) \in D}[e_k^v(x)] \approx \mathbb{E}_{(x, y=k') \in D}[e_{k'}^v(x)] \qquad \forall k,k' \in \{1,\dots,K\}.$$3 handles class imbalance and that a coefficient $$\mathbb{E}_{(x, y=k) \in D}[e_k^v(x)] \approx \mathbb{E}_{(x, y=k') \in D}[e_{k'}^v(x)] \qquad \forall k,k' \in \{1,\dots,K\}.$$4 balances the consistency term. Optimization uses Adam with L2 regularization, runs for 200 epochs, activates the adaptive prior after 20 epochs, updates the prior every 5 epochs, and repeats experiments with 5 random seeds (Chen et al., 18 Aug 2025).

The evaluation is conducted on five real-world multi-view datasets: Handwritten, Animal, Scene15, YaleB, and Caltech-101. The reported baselines include the single-view evidential methods TLC, I-EDL, and R-EDL, as well as the multi-view evidential methods TMC, CCML, and ECML. Performance is reported using ACC and ECE.

The main numerical results presented in the source are as follows. On Handwritten, FAML reaches 94.2% ACC, compared with the best baseline around 85.5%. On Animal, it reaches 76.2%, compared with baselines around 68–69%. On Scene15, it reaches 57.2%, compared with approximately 44% for the best baseline. On YaleB, it reaches 90.4%, compared with approximately 83.7%. On Caltech-101, it reaches 87.1%, compared with approximately 76%. The paper also reports that ECE is consistently lower for FAML, supporting the claim that its uncertainty estimation is more reliable.

The visualization-based findings are equally central to the BEML argument. Baseline methods such as TMC are described as assigning low uncertainty to head classes but high uncertainty to tail classes, and as allocating much more evidence to data-rich classes. FAML, by contrast, is reported to produce more uniform evidence strength across classes and more consistent uncertainty. This is precisely the empirical signature that the BEML formulation is intended to correct.

An ablation on Caltech-101 isolates the contributions of the adaptive prior, fairness constraint, and consistency regularizer:

Configuration	ACC
No adaptive prior, no fairness, no consistency	73.8
Adaptive prior only	79.1
Adaptive prior + fairness	84.3
Adaptive prior + consistency	80.0
Adaptive prior + fairness + consistency	87.1

The reported interpretation is that adaptive prior alone yields a clear improvement over the plain baseline, the fairness constraint contributes more than consistency alone, and all three components together achieve the best result. Sensitivity analysis varies the prior-strength parameter $$\mathbb{E}_{(x, y=k) \in D}[e_k^v(x)] \approx \mathbb{E}_{(x, y=k') \in D}[e_{k'}^v(x)] \qquad \forall k,k' \in \{1,\dots,K\}.$$5 from 0.1 to 10.0 and reports that performance is generally best when $$\mathbb{E}_{(x, y=k) \in D}[e_k^v(x)] \approx \mathbb{E}_{(x, y=k') \in D}[e_{k'}^v(x)] \qquad \forall k,k' \in \{1,\dots,K\}.$$6 lies in [1.0, 10.0].

6. Relation to adjacent evidential multi-view research

BEML is closely related to, but distinct from, other evidential multi-view formulations. "Reliable Conflictive Multi-View Learning" introduces the RCML problem and the corresponding ECML method for settings in which views may be noisy, unaligned, or otherwise conflictive across views (Xu et al., 2024). ECML also learns view-specific evidence, converts it to Dirichlet opinions, and attaches uncertainty and reliability to predictions, but its central concern is how fused uncertainty should respond when views disagree. In that framework, the defining question is whether conflictive instances should yield higher uncertainty rather than being removed or repaired. By contrast, BEML identifies a different pathology: biased evidence generation within each view, especially toward data-rich classes. The distinction matters because a model can be conflict-aware without being fair in its class-wise evidence allocation.

A second adjacent line is represented by "MERIT: Multi-view evidential learning for reliable and interpretable liver fibrosis staging" (Liu et al., 2024). MERIT also uses subjective-logic opinions and introduces a distribution-aware base rate,

$$\mathbb{E}_{(x, y=k) \in D}[e_k^v(x)] \approx \mathbb{E}_{(x, y=k') \in D}[e_{k'}^v(x)] \qquad \forall k,k' \in \{1,\dots,K\}.$$7

to account for class distribution effects and possible class distribution shift. Its fusion strategy combines local views with cumulative belief fusion and then combines local and global opinions with belief constraint fusion. This suggests a conceptual proximity to BEML insofar as non-uniform priors or base rates are used to mitigate class-distribution effects. However, the BEML formulation in FAML is more explicit about fairness in evidence allocation: it defines a fairness target over expected ground-truth evidence, adds a class-wise evidence variance constraint, and ties the prior to the training trajectory rather than to a fixed empirical class prior.

Taken together, these works delimit three nearby but non-identical problem classes in evidential multi-view learning. RCML/ECML focuses on conflictive views and reliability-aware fusion; MERIT emphasizes reliability, interpretability, and distribution-aware base rates in a domain-specific clinical setting; BEML/FAML focuses on class-frequency-induced evidence bias and the resulting failure of uncertainty trustworthiness. A plausible implication is that future evidential multi-view systems may need to account for all three phenomena simultaneously: biased evidence generation, conflictive inter-view structure, and prior or base-rate mismatch under distribution shift.