Fairness-Aware Multi-View Evidential Learning (FAML)

• The paper introduces an adaptive training-trajectory-based prior and explicit fairness regularization to correct biased evidence allocation across imbalanced classes.
• It employs confidence-weighted opinion alignment to fuse view-specific evidential predictions, ensuring balanced uncertainty and reliable multi-view integration.
• Empirical results on diverse datasets demonstrate improved overall accuracy and uncertainty calibration, particularly enhancing performance on minority classes.

Searching arXiv for the cited FAML paper and closely related multi-view evidential/fairness work. arXiv search query: (Chen et al., 18 Aug 2025) Fairness-Aware Multi-view Evidential Learning with Adaptive Prior Fairness-Aware Multi-view Evidential Learning (FAML) is a multi-view evidential learning framework that addresses what its authors term the Biased Evidential Multi-view Learning (BEML) problem: the tendency of view-specific evidential models to assign disproportionately strong evidence to data-rich classes, thereby skewing both prediction and uncertainty estimation (Chen et al., 18 Aug 2025). In FAML, “fairness” denotes balanced evidence allocation across classes rather than demographic fairness over sensitive groups. The method combines view-specific evidential prediction, an adaptive prior derived from training trajectory, an explicit fairness constraint based on class-wise evidence variance, and an opinion alignment mechanism for multi-view fusion, with the stated goal of improving both classification performance and the reliability of uncertainty estimation (Chen et al., 18 Aug 2025).

1. Problem formulation and the meaning of fairness

FAML considers supervised multi-view classification with dataset

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

where $x_n^v \in \mathbb{R}^{d_v}$ is the feature vector of instance $n$ in view $v$, $y_n \in \{1,\dots,K\}$ is the class label, $V$ is the number of views, and $K$ is the number of classes (Chen et al., 18 Aug 2025). The target setting is explicitly imbalanced: training data are constructed to contain head, medium, and tail regions by sampling with a Pareto distribution, while test sets remain balanced.

The motivating observation is that prior multi-view evidential learning commonly assumes that view-specific evidence learning is already reliable and concentrates methodological effort on the fusion stage. FAML challenges that premise by arguing that evidence generation itself is systematically biased on imbalanced real-world data. The paper states that samples tend to be assigned more evidence to support data-rich classes, minority or tail classes tend to receive less supportive evidence even when correctly classified, and uncertainty therefore becomes unreliable: some wrong predictions toward head classes become overconfident, while some correct predictions for tail classes become underconfident (Chen et al., 18 Aug 2025).

This is formalized as the BEML problem. The paper states that, for fair evidence allocation, the expected evidence assigned to the ground-truth class should be comparable across classes in each view: $$\mathbb{E}_{(x, y=k) \in D}\left[e_k^v(x)\right] \approx \mathbb{E}_{(x, y=k') \in D}\left[e_{k'}^v(x)\right], \quad \forall k,k' \in \{1,\dots,K\}.$$ Accordingly, the method’s notion of fairness is class-wise: it seeks to prevent the model from systematically assigning stronger supportive evidence to head classes merely because they are data-rich (Chen et al., 18 Aug 2025).

A persistent source of confusion is the word “fairness.” In FAML, fairness is not defined over protected groups, demographic parity, or equalized odds. It is defined as balanced evidence support across classes. This makes FAML a fairness method in the sense of evidence allocation and calibration under class imbalance, not in the sense used by demographic fairness literature.

2. Evidential and subjective-logic foundations

FAML is built on the standard evidential deep learning pipeline in which each view-specific network outputs a nonnegative evidence vector

$\mathbf{e} = \{e_1,\dots,e_K\}, \qquad e_k \ge 0.$

Under the standard mapping, Dirichlet concentration parameters are

$\alpha_k = e_k + 1,$

with Dirichlet strength

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

Subjective-logic belief masses and uncertainty are then

$x_n^v \in \mathbb{R}^{d_v}$1

and the projected class probability is

$x_n^v \in \mathbb{R}^{d_v}$2

where $x_n^v \in \mathbb{R}^{d_v}$3 is the base rate (Chen et al., 18 Aug 2025).

These quantities have the usual evidential interpretation. Larger evidence increases $x_n^v \in \mathbb{R}^{d_v}$4, which lowers $x_n^v \in \mathbb{R}^{d_v}$5; low evidence implies high uncertainty. A view-specific prediction is represented as a multinomial opinion

$x_n^v \in \mathbb{R}^{d_v}$6

Confidence is therefore not treated as a separate learned scalar, but is induced by the interaction between class probability, total evidence strength, and residual uncertainty.

The supervised evidential loss used by FAML is the expected cross-entropy under the Dirichlet posterior. For sample $x_n^v \in \mathbb{R}^{d_v}$7,

$x_n^v \in \mathbb{R}^{d_v}$8

where $x_n^v \in \mathbb{R}^{d_v}$9, $n$0 is one-hot, $n$1 is the multivariate Beta function, and $n$2 is the digamma function (Chen et al., 18 Aug 2025).

This construction situates FAML squarely within evidential multi-view learning rather than ordinary probabilistic ensembling. The method’s novelty is not the Dirichlet machinery itself, but the claim that the evidence-generation process before fusion is biased and requires explicit correction.

3. Adaptive prior and explicit fairness regularization

FAML departs from standard evidential deep learning by replacing the fixed uniform prior with an adaptive class-specific prior based on training trajectory. The paper defines the correctness indicator

$n$3

For class $n$4, the adaptive prior is

$n$5

where $n$6 is the number of training samples in class $n$7 and $n$8 controls prior strength (Chen et al., 18 Aug 2025). The Dirichlet concentration is then modified to

$n$9

The intended effect is corrective. When a class has fewer correct predictions relative to its sample count, $v$0 becomes larger, providing additional prior support to an underperforming class. Conversely, classes already being learned well receive less correction. The paper states two propositions about this mechanism. First, the adaptive training-trajectory-based prior is a generalized form of the uniform prior employed by standard evidential deep learning and progressively converges to the uniform prior under ideal training conditions. Second, the adaptive prior implicitly reformulates the base rate in the projected probability of a subjective opinion, increasing predictive probability mass for underperforming classes through biased uncertainty allocation (Chen et al., 18 Aug 2025).

FAML complements this implicit calibration with an explicit fairness term. The paper defines the Fairness Degree as the variance of class-wise average evidence: $v$1 where $v$2 is the average evidence for class $v$3 and $v$4 is the global average over class-wise evidence means (Chen et al., 18 Aug 2025). High variance indicates that some classes systematically receive much more evidence than others; minimizing this term encourages a more balanced evidence profile.

The fairness-augmented classification loss is

$v$5

The paper states that $v$6 is gradually changed from $v$7 to $v$8 during training, so the fairness constraint is introduced progressively rather than dominating the earliest optimization phase (Chen et al., 18 Aug 2025).

This pairing of adaptive prior and evidence-variance regularization is the conceptual center of FAML. One corrects class imbalance through prior adjustment; the other directly constrains disparity in evidence allocation. Taken together, they shift the fairness question from output labels to the internal evidential substrate.

4. Opinion alignment and multi-view fusion

FAML argues that evidence bias is not only class-dependent but also view-specific. One view may incorrectly favor one data-rich class, another view a different one. To mitigate this, the method introduces an opinion alignment mechanism before or during fusion (Chen et al., 18 Aug 2025).

For two views $v$9 and $y_n \in \{1,\dots,K\}$0, the paper defines confidence-weighted evidence aggregation as

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

where

$y_n \in \{1,\dots,K\}$2

Thus more certain views contribute more strongly to the aggregated evidence (Chen et al., 18 Aug 2025).

To measure disagreement across views, FAML uses the variance of the Dirichlet. For class $y_n \in \{1,\dots,K\}$3 in view $y_n \in \{1,\dots,K\}$4,

$y_n \in \{1,\dots,K\}$5

The dissonance degree between two views is then

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

and the consistency loss over all views is

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

Minimizing this term encourages views to maintain more consistent uncertainty-aware opinions, reducing the chance that one confidently biased view will dominate the fused decision (Chen et al., 18 Aug 2025).

The overall loss is class-balanced: $y_n \in \{1,\dots,K\}$8 where $y_n \in \{1,\dots,K\}$9 is the number of samples in the class of sample $V$0’s label, $V$1 weights opinion consistency, and both fused and per-view outputs are supervised (Chen et al., 18 Aug 2025).

Although the paper references Subjective Logic and trusted multi-view evidence aggregation, the fusion actually used in FAML is the above confidence-weighted evidence combination rather than an explicit Dempster–Shafer orthogonal sum. This distinction matters because FAML’s contribution lies less in a new fusion algebra than in bias-aware evidence generation and alignment.

5. Optimization protocol and empirical behavior

FAML uses view-specific evidential networks implemented as fully connected networks with a ReLU layer, trained in PyTorch 1.13.1 with Adam, $V$2-norm regularization, and 200 epochs; results are reported over 5 random seeds as mean $V$3 standard deviation (Chen et al., 18 Aug 2025). The adaptive prior is not updated in the earliest phase. The paper states that prior adjustment starts after 20 epochs and is refreshed every 5 epochs thereafter, because early predictions are too unstable to support reliable class-wise calibration.

The experimental benchmark contains five real-world multi-view datasets: Handwritten, Animal, Scene15, YaleB, and Caltech-101 (Chen et al., 18 Aug 2025). Baselines include single-view evidential methods—TLC, I-EDL, and R-EDL—and multi-view evidential methods—TMC, CCML, and ECML. Performance is reported with ACC and ECE, together with Head, Med, and Tail splits.

Dataset	ACC	ECE
Handwritten	$V$4	$V$5
Animal	$V$6	$V$7
Scene15	$V$8	$V$9
YaleB	$K$0	$K$1
Caltech-101	$K$2	$K$3

The paper reports that FAML achieves the best overall ACC on all five datasets and the lowest overall ECE on every dataset (Chen et al., 18 Aug 2025). Tail-class results are especially emphasized: Handwritten tail $K$4, Animal tail $K$5, Scene15 tail $K$6, YaleB tail $K$7, and Caltech-101 tail $K$8. The empirical claim is therefore not only higher overall accuracy, but stronger behavior precisely where evidence bias is expected to be most severe.

The ablation on Caltech-101 isolates the main components. The configuration with no adaptive prior, no fairness, and no consistency attains $K$9; adaptive prior only reaches $$\mathbb{E}_{(x, y=k) \in D}\left[e_k^v(x)\right] \approx \mathbb{E}_{(x, y=k') \in D}\left[e_{k'}^v(x)\right], \quad \forall k,k' \in \{1,\dots,K\}.$$0; adaptive prior plus fairness reaches $$\mathbb{E}_{(x, y=k) \in D}\left[e_k^v(x)\right] \approx \mathbb{E}_{(x, y=k') \in D}\left[e_{k'}^v(x)\right], \quad \forall k,k' \in \{1,\dots,K\}.$$1; adaptive prior plus consistency reaches $$\mathbb{E}_{(x, y=k) \in D}\left[e_k^v(x)\right] \approx \mathbb{E}_{(x, y=k') \in D}\left[e_{k'}^v(x)\right], \quad \forall k,k' \in \{1,\dots,K\}.$$2; and the full model reaches $$\mathbb{E}_{(x, y=k) \in D}\left[e_k^v(x)\right] \approx \mathbb{E}_{(x, y=k') \in D}\left[e_{k'}^v(x)\right], \quad \forall k,k' \in \{1,\dots,K\}.$$3 (Chen et al., 18 Aug 2025). These numbers support the paper’s interpretation that all three ingredients matter, with the fairness constraint contributing more than the consistency term in this setting.

Qualitative analyses reinforce the same point. On Animal, TMC is reported to yield low uncertainty for head classes and much higher uncertainty for medium and tail classes, whereas FAML produces more uniform and lower uncertainty. On Handwritten, TMC is reported to collect much more evidence for data-rich classes, whereas FAML yields more balanced average evidence across categories (Chen et al., 18 Aug 2025). This suggests that FAML improves not just output calibration but the internal distribution of evidential support.

6. Relation to adjacent research and broader interpretation

FAML belongs to a line of trusted multi-view evidential methods initiated by work such as TMC, which introduced per-view evidence extraction, Dirichlet opinions, and uncertainty-aware fusion through subjective logic and Dempster–Shafer style combination (Han et al., 2022). Later work such as ECML emphasized conflictive multi-view learning and fusion rules that can increase or decrease uncertainty depending on incoming view reliability (Xu et al., 2024), while TMUR argued that independently estimated view uncertainties are not necessarily numerically comparable and proposed unified global routing instead of branch-local arbitration (Zhang et al., 10 Apr 2026). Relative to these systems, FAML’s central claim is different: the evidence-generation stage itself is biased and requires correction before fusion (Chen et al., 18 Aug 2025).

The paper also sits near, but not inside, the demographic fairness literature on multi-view learning. FairMVC injects group fairness into soft cluster assignment so that the fraction of different groups in each cluster is approximately identical to the entire data set, and AFMVC moves fairness intervention into representation learning through adversarial removal of sensitive information from latent features (Zheng et al., 2023, Jiang et al., 6 Aug 2025). Those methods study group fairness over sensitive attributes, whereas FAML studies class-wise fairness in evidence support. The shared vocabulary can be misleading, but the underlying objectives differ substantially.

A plausible implication is that FAML addresses a calibration and allocation problem that many evidential multi-view systems leave implicit. The method does not ask whether uncertainty is merely available; it asks whether evidence is being learned equitably across classes before uncertainty is computed. This suggests a broader reinterpretation of evidential learning: trustworthiness may depend as much on how evidence is distributed across categories as on the formal correctness of the fusion operator.

At the same time, FAML has clear scope conditions. Its fairness notion is class-level only. It does not model protected attributes, subgroup disparities, demographic parity, or equalized odds (Chen et al., 18 Aug 2025). The adaptive prior relies on training correctness, so noisy labels or unstable early optimization can affect prior estimates. The variance-based fairness term may also be coarse when intrinsic class difficulty differs substantially. These are not contradictions of the method’s stated goals, but they delimit its meaning.

Within multi-view evidential learning, FAML’s enduring contribution is the reframing of the central question. Earlier work often asked how to fuse evidence under noise, conflict, incompleteness, or incomparable scales (Han et al., 2022, Xu et al., 2024, Zhang et al., 10 Apr 2026). FAML adds a prior question: is the evidence itself being learned fairly? (Chen et al., 18 Aug 2025)