Computer scientists evaluating automated emotion recognition systems have introduced a multi-strategy enhanced dung beetle optimizer that significantly improves both predictive accuracy and probabilistic reliability during hyperparameter tuning. Tested across four controlled feature domains derived from the Multimodal EmotionLines Dataset, the new optimization algorithm achieved an average test weighted F1 score of 79.85 percent, outperforming the standard dung beetle optimizer by 4.63 percentage points while consistently reducing run-to-run variance across independent experimental trials. Beyond classification performance, the optimization framework simultaneously prioritized model trustworthiness by obtaining the lowest multiclass Brier score of 0.146 and the lowest negative log-likelihood of 0.425 among seven competing search strategies, including random sampling, differential evolution, and advanced Bayesian methods. When paired with an automated confidence-based deferral mechanism at a decision threshold of 0.55, the selected neural network configurations achieved a selective weighted F1 score of 85.31 percent while retaining 84.8 percent of all predictions, demonstrating how joint optimization of calibration and accuracy creates dependable affective computing architectures under strict computational limits.
Automated multimodal emotion recognition integrates linguistic, acoustic, and visual information to identify emotional expressions across conversational exchanges, but configuring the underlying neural networks remains notoriously difficult. In real-world environments, raw data streams are frequently contaminated by background noise, recording discrepancies, and temporary signal loss, which causes conventional validation metrics to fluctuate erratically during model selection. Standard machine learning workflows that seek only to maximize top-line validation accuracy routinely select hyperparameter configurations that suffer from poor probability calibration, severe sensitivity to missing inputs, and inflated computational demands. A model that displays high predictive accuracy in training may issue dangerously overconfident judgments when speech or facial features become corrupted or absent. Addressing this fundamental reliability gap requires an optimization framework that evaluates candidate architectures against a composite objective balancing predictive power, probability fidelity, and inference efficiency, while methodically exploring mixed discrete and continuous search spaces under a strictly enforced budget of objective function evaluations.
To overcome these challenges, the researchers modified the biological heuristics of the dung beetle optimizer, an evolutionary algorithm inspired by the distinct rolling, breeding, foraging, and stealing movement behaviors observed in natural populations. Although standard dung beetle optimization provides complementary exploration and exploitation operators, its static allocation of population roles prevents candidate solutions from adapting dynamically to changing population diversity, making the optimizer vulnerable to premature convergence in irregular validation landscapes. The newly developed algorithm, designated MSA-DBO, addresses these weaknesses through four integrated enhancements: opposition-based initialization, diversity- and remaining-budget-dependent behavioral-role allocation, stagnation-triggered bounded chaotic escape, and an explicit feasibility-repair operator. Working within a normalized eight-dimensional search space, the optimizer initializes an active population of six candidates alongside six coordinate-wise opposition vectors, thereby maximizing initial search-space coverage across continuous and discrete bounds before candidate evaluation begins without exceeding a preallocated total budget of thirty full-fidelity candidate evaluations per search run.
The algorithm monitors the concentration of candidate solutions across the normalized search space by computing a coordinate-dispersion index at each generational update. When latent dispersion drops below a predefined threshold of 0.15 while more than 40 percent of the evaluation budget remains, the optimizer dynamically alters its behavioral-role distribution from the default configuration of one roller, two breeders, one forager, and two stealers to an exploratory profile consisting of two rollers, one breeder, two foragers, and one stealer. This shift broadens population movement before local convergence sets in. If the archive-leading global best solution fails to improve over two consecutive generational updates, the optimizer triggers a bounded chaotic escape mechanism based on a logistic map with a control parameter of four. The magnitude of this chaotic perturbation scales down proportionally with the remaining evaluation budget from an initial maximum coefficient of 0.20, enabling substantial exploratory jumps early in the search while preserving fine-grained local refinement near optimization termination.
Every candidate vector generated by the optimizer represents a complete set of architectural and training hyperparameters that define a two-hidden-layer multilayer perceptron. The encoded dimensions determine the width of the first hidden layer from 32, 64, 96, or 128 units, the width of the second hidden layer from 16, 32, 48, or 64 units, and the activation function from rectified linear, exponential linear, or hyperbolic tangent activations. Continuous parameters control the L2 weight regularization coefficient across a logarithmic range from 10^-6 to 10^-2, the Adam learning rate across 10^-4 to 5*10^-3, and feature dropout from zero to 0.50. The remaining two coordinates specify text and acoustic modality logits relative to a fixed visual baseline of zero, producing global softmax weights that govern input fusion.
During candidate evaluation, each proposed neural architecture is trained on a dedicated training subset using categorical cross-entropy loss, mini-batches of 32 samples, and early stopping after seven stagnant epochs up to a maximum ceiling of 50 epochs. Candidate selection is driven by an uncertainty-aware validation function combining weighted F1 error with an importance weight of 0.70, expected calibration error with a weight of 0.10, multiclass negative log-likelihood with a weight of 0.15, and normalized model fitting time with a weight of 0.05. By explicitly penalizing overconfident classification errors and computational latency alongside misclassifications, the composite validation metric prevents the selection of oversized, poorly calibrated architectures. Once the thirty-candidate search budget is exhausted, the configuration maintaining the lowest validation score across an elitist archive is evaluated on held-out test data, with predictions categorized into high-confidence, moderate-confidence, and low-confidence abstention bins based on calibrated maximum softmax probabilities.
The experimental evaluation utilized 2,400 dialogue utterances from the Multimodal EmotionLines Dataset, divided equally among four non-overlapping controlled feature domains representing diverse operational stresses: dialogue-noisy, naturalistic-imbalanced, balanced-reference, and challenge-shifted environments. Each domain was partitioned using a strict dialogue-group-preserving 60/20/20 train, validation, and test split, guaranteeing that individual conversational contexts never overlapped across evaluation partitions. In the balanced-reference domain, Gaussian-process Bayesian optimization achieved a marginally higher score of 81.08 percent compared to 80.50 percent for MSA-DBO. Across all domains, the enhanced optimizer demonstrated substantial gains over standard dung beetle optimization, boosting weighted F1 by 6.49 percentage points in noisy conditions and 6.30 percentage points under challenge-shifted distributions.
To rigorously verify whether performance advantages were statistically significant across the twenty matched domain-seed replications, the researchers conducted two-sided Wilcoxon signed-rank tests adjusted for multiple testing using the Holm procedure. The aggregate comparisons demonstrated statistically significant advantages for MSA-DBO over standard dung beetle optimization, with a mean difference of 4.63 percentage points, an adjusted p-value below 0.001, and a paired rank-biserial correlation of 0.981. Paired comparisons against random search, differential evolution, Bayesian Optimization and Hyperband, and Optuna-TPE also yielded statistically significant differences, with adjusted p-values of 0.001, 0.002, 0.011, and 0.011, respectively. However, the aggregate comparison against Gaussian-process Bayesian optimization revealed a smaller mean difference of 1.37 percentage points and an adjusted p-value of 0.058, falling short of statistical significance at the five percent level. When analyzed within individual domain subsets, performance differences among leading adaptive optimizers were statistically indistinguishable, showing that aggregate gains reflect consistent multi-domain robustness rather than uniform domain-specific dominance.
Systematic stress testing under controlled test-time modality masking highlighted how hyperparameter choices govern resilience when sensor inputs fail completely. When text, acoustic, or visual feature blocks were individually set to zero post-selection without model retraining, MSA-DBO maintained the highest weighted F1 across all single-omission scenarios, suffering its largest drop of 7.75 percentage points when linguistic information was removed. Under severe pairwise omissions where two data streams were simultaneously eliminated, Bayesian Optimization and Hyperband achieved a slightly higher average score of 67.10 percent versus 66.85 percent for MSA-DBO, reflecting the front-loaded resource efficiency of multi-fidelity pruning. Component ablation experiments revealed that removing the uncertainty-aware objective caused the steepest drop in weighted F1, reducing performance by 3.11 percentage points, followed by opposition initialization, adaptive role allocation, feasibility repair, and chaotic escape. Computationally, MSA-DBO incurred only a 2.6 percent increase in total wall-clock search time over standard DBO while selecting more compact networks averaging 2,462 parameters.
Exploratory linear regression diagnostics examining the twenty validation-selected MSA-DBO configurations showed that a smaller validation-to-test generalization gap was associated with superior test confidence quality, with a standardized regression coefficient of negative 0.44 and an exact p-value of 0.021. The authors emphasized that positive associations between confidence quality and decision utility are partly mathematical artifacts arising from shared calibration terms rather than direct empirical proof of autonomous decision efficacy. Furthermore, the researchers noted essential operational boundaries: all experiments operated on pre-extracted feature representations rather than end-to-end multimodal neural pipelines, test-time masking did not simulate asynchronous sensor dropouts or non-random missingness, and metaheuristic operators lack formal mathematical convergence proofs. Future investigations will evaluate adaptive role allocation and chaotic perturbation across independent affective benchmarks, raw audio-visual streams, and unconstrained conversational settings to establish whether these algorithmic reliability mechanisms extend to large-scale deep learning architectures.
Subject of Research: Reliability-oriented hyperparameter optimization for multimodal emotion recognition using an enhanced metaheuristic algorithm
Article Title: Multi-strategy enhanced dung beetle optimizer for reliable hyperparameter search in multimodal emotion recognition
Article References: Rui, Y. (2026). Multi-strategy enhanced dung beetle optimizer for reliable hyperparameter search in multimodal emotion recognition. Discover Computing, 29(1), Article 797. https://doi.org/10.1007/s10791-026-10777-8
Image Credits: AI Generated
DOI: 10.1007/s10791-026-10777-8
Keywords: multimodal emotion recognition, hyperparameter optimization, dung beetle optimizer, metaheuristic algorithms, expected calibration error, Brier score, negative log-likelihood, selective prediction, confidence deferral, opposition-based initialization, chaotic escape, feasibility repair, MELD benchmark, modality masking, model selection, computational complexity, population diversity
News Source: Denise Maddox. (November 12, 2026). Enhanced Dung Beetle Optimizer Boosts Reliability in Multimodal Emotion Recognition Models. Scienmag.
