Multi-View Voting: Theory and Applications

• Multi-view voting is an ensemble method that integrates diverse predictions from distinct data views, balancing accuracy, diversity, and robustness against noise.
• It leverages theoretical frameworks like PAC-Bayesian analysis and optimized weighting strategies to achieve strong generalization and reliable decision fusion.
• Applied in fields such as medical imaging, remote sensing, and social choice, multi-view voting improves data fusion and mitigates noise-induced errors.

Multi-view voting is a foundational ensemble strategy in which predictions, judgments, or signals obtained from multiple distinct "views" or data representations are aggregated using a voting-based integration scheme. This paradigm is central to multi-view learning, machine ensemble methods, modern computational social choice, and high-reliability decision-fusion in domains including classification, outlier detection, medical imaging, and election control. The theoretical and algorithmic developments in multi-view voting address both the optimal aggregation of heterogeneous or complementary sources, and the trade-offs between accuracy, diversity, and robustness against noise, bias, or adversarial manipulation.

1. Theoretical Foundations: Multi-View Majority Vote and PAC-Bayesian Analysis

In statistical learning, a multi-view majority vote classifier is defined on data $x=(x^{(1)},...,x^{(V)})$ where each component corresponds to a distinct view or representation. For each view $v$, a distribution $\mathcal{Q}^v$ over base classifiers (voters) $h^v:\mathcal{X}^v \to \mathcal{Y}$ is specified, together with a hyper-posterior $\rho$ over the views. The final aggregated prediction is given by

$$B_\rho(x) = \arg\max_{y\in\mathcal{Y}} \mathbb{E}_{v\sim\rho}\mathbb{E}_{h\sim\mathcal{Q}^v}[\mathbb{I}\{h(x^v) = y\}],$$

which is typically only tractable to optimize or analyze via surrogate risk, margin bounds, or relaxation to randomized voting (Gibbs) predictors. PAC-Bayesian theory extends to the multi-view setting by leveraging Rényi divergences between posteriors and priors at both the voter and the view level, yielding generalization bounds such as

$$\mathrm{KL}(\hat{\mathfrak{R}}_S^{\mathcal{V}}\| \mathfrak{R}_{\mathcal{D}}^{\mathcal{V}}) \leq \frac{1}{m}\left(\mathbb{E}_{\rho}[D_\alpha(\mathcal{Q}^v \|\mathcal{P}^v)]+ D_\alpha(\rho\|\pi) + \log \frac{2\sqrt{m}}{\delta}\right),$$

enabling tight probabilistic control of ensemble error in the multi-view context (Hennequin et al., 2024). These bounds generalize classic single-view majority vote theorems and directly account for the statistical structure imposed by the view hierarchy.

2. Multi-View Voting in Ensemble and Boosted Learning

Multi-view voting is operationalized in a range of methods:

• PB-MVBoost (Goyal et al., 2018): A boosting-style algorithm where, at each round, weak learners are trained per view, and view weights $\rho$ are adaptively optimized to maximize a PAC-Bayesian C-bound that exploits the dual desiderata of minimizing the Gibbs risk (average error) and maximizing the inter-view disagreement (diversity). The multi-view C-bound

$$R_\mathcal{D}(B_\rho) \leq 1 - \frac{(1 - 2 R_\rho)^2}{1 - 2 \mathrm{Dis}_\rho}$$

quantifies how both accuracy and diversity play critical roles in ensuring strong generalization, and PB-MVBoost directly navigates this tradeoff via joint optimization.

• Multi-view Bregman Boost (Goyal et al., 2018): Here, both within-view and across-view majority votes are learned, with weights optimized through Bregman divergence minimization. Mirror-descent updates jointly adjust per-view and view-level weights in the simplex, minimizing a convex surrogate risk and leading to fast, parallelizable global optimum search in high-dimensional multi-view spaces.
• Dynamic and Personalized Voting (Cao et al., 2018): In radiomics and scenarios with patient-specific heterogeneity, per-example dynamic voting weights aggregate the outputs of Random Forests trained on different feature views, leveraging both global model confidence and locality-adaptive competence (e.g., out-of-bag accuracy on similar neighbors), surpassing static and majority voting in experimental accuracy.

3. Applications in Structured Data Fusion and Robust Aggregation

Multi-view voting underpins robust data fusion in diverse applied contexts:

• Medical Image Segmentation (Ding et al., 2020): Multi-view CNNs trained on orthogonal anatomical planes generate per-view segmentations, which are fused via channel-wise majority vote or (weighted) averaging at the voxel level. Voting-based fusion is effective for integrating spatially heterogeneous evidence and is complemented by multi-view fusion loss functions for joint training.
• Remote Sensing Label Purification (Wang et al., 2023): To filter out label noise, data are partitioned into disjoint "views," with independent models trained and their predictions aggregated through unanimity voting. Instances for which all views agree are deemed reliable, with high-entropy/ambiguous cases iteratively relabeled using an additional model, yielding substantial increases in robustness to label corruption.
• Provenance-based Intrusion Detection (Yang et al., 16 Apr 2026): ProvFusion computes heterogeneous anomaly scores from attribute, structural, and causal perspectives for every node in a provenance graph, fuses the normalized scores through a family of monotonic detectors, and applies a voting threshold to aggregate signals. This pipeline extends the coverage of detection to both node-centric and edge-centric anomalies, demonstrating lower false positive rates and stronger generalization across security benchmarks.

4. Multi-View Voting in Social Choice and Election Control

The multi-view (or multi-vote) paradigm appears in computational social choice, where each "view" may correspond to separate aspects or issues in elections. In the control-by-selecting-rules problem (Wang et al., 2023), each voter casts a ballot for each view/layer, which is scored under a rule mapping, and aggregated via sum, max, or min to determine satisfaction. The complexity of controlling outcomes via selective rule assignment is highly intractable (NP-hard, W[1]/W[2]-hard under various parameterizations), even in highly restricted cases, delineating algorithmic boundaries for election manipulation and underscoring the resistance of multi-view elections to procedural control.

In multi-issue approval voting (Mazur et al., 2017), voters approve platforms in a product space (e.g., time × location), and the fraction of voters who approve any common platform is provably lower than in the single-issue case, following sharp combinatorial and topological bounds. This demonstrates, for product societies, that the guaranteed agreement fraction is the product or minimum of one-dimensional bounds, showing the fundamental limiting effect of multi-view aggregation on group consensus.

5. Optimization Algorithms and Empirical Performance

The direct minimization of PAC-Bayesian bounds for multi-view majority voting (Hennequin et al., 2024) is operationalized by self-bounding gradient-based algorithms (both for PAC-Bayes-λ and inverted-KL formulations). These optimize voter and view posteriors jointly, enforce divergences via log-barrier penalties, and are computationally tractable (using samples or efficient bisection where closed-form gradients are unavailable). Empirical evaluations across 10 multi-view datasets show that optimal multi-view voting generally outperforms single-view baselines and concatenation, with particular gains in robustness to view poisoning and structured interpretability via learned weights.

6. Impact, Limitations, and Future Directions

Multi-view voting integrates statistical learning, ensemble design, robust data fusion, and collective decision-making under a unified theoretical and applied umbrella. Recent developments quantify trade-offs between expressivity, diversity, and computational hardness, providing principled guarantees and practical algorithms for complex structured domains.

Limitations observed include:

• Increased complexity of joint optimization, particularly as the number of views or the base voter pools scale.
• Situational necessity for tuning surrogate loss/exponent parameters for best empirical performance.
• The combinatorial explosion of possibilities in rule selection and decision aggregation in the social choice context.

Future research directions include semi-supervised and transfer extensions (incorporating unlabeled data), structured output generalizations, adaptive fusion weights (including learnable or attention-based gating), and new statistical guarantees leveraging non-KL divergence measures (Hennequin et al., 2024).

In sum, multi-view voting frameworks systematize and enable principled exploitation of complementary, heterogeneous, or redundant information sources, returning decisions that are more robust, interpretable, and theoretically grounded than single-view or naïvely fused alternatives.