Journal of Computing Theories and Applications

CAMMRF: A Conflict-Aware Missing-Modality Robust Fusion Framework for Multimodal Sentiment Analysis under Heterogeneous Observability

Kajal Chourasiya — 2026-09-11

Multimodal sentiment analysis combines linguistic, acoustic, and visual evidence, yet deployment commonly violates the benchmark assumption that every modality is present and mutually consistent. Missing-modality systems generally reconstruct absent channels while conflict-aware systems generally assume full observation, even though absence and disagreement both change how far an observed modality should be trusted. We therefore treat the two jointly and hypothesise that conditioning modality reliability on both availability and cross-modal compatibility, and applying it across aligned and disagreement-preserving representations, yields more consistent performance across observability regimes than matched fusion controls without this coupling. We introduce CAMMRF, a framework that couples a mask- and peer-conditioned reliability estimator, an alignment/conflict subspace decomposition, and stochastic modality dropout with cross-view consistency. We evaluate one frozen protocol on the official CMU-MOSI train/validation/test partitions under eight observability regimes; five independent training seeds (42, 43, 44, 45, and 46) completed, each retaining predictions, checkpoints, and logs. In the full-modality regime CAMMRF is best on all four metrics simultaneously—MAE 1.050 ± 0.028, Pearson correlation 0.593 ± 0.009, weighted F1 75.01 ± 1.72%, and Acc-2 74.91 ± 1.80%—rather than on accuracy alone, and it ranks first by mean Acc-2 in 5 of eight regimes. The advantage persists when audio or vision is missing and under the controlled conflict stress (mean Acc-2 73.81%, a 1.10-point decrease, with matching small changes in MAE, weighted F1, and correlation); however, all four metrics degrade together when text is absent—correlation most steeply—so robustness is regime-dependent rather than universal. Full-budget ablations show that several component removals improve full-observation point estimates, supporting an accuracy–robustness trade-off rather than an independent-component claim. The evidence is limited to one corpus and controlled surrogate baselines and does not establish external generalisation. A single Colab notebook regenerates the protocol and exports a machine-readable result registry.

Exhaustive Entanglement-Topology Enumeration and Edge-Vertex-Level Attribution in Quantum Feature Map Design for Classification

Natanael Natanael — 2026-09-11

Quantum kernel methods encode classical data into quantum states using specially designed feature-map circuits and then train a classical SVM on the resulting kernel matrix. Prior studies of entanglement structure typically compare only a few basic topologies, such as linear, circular, and full entanglement, leaving the broader space of entanglement graphs largely unexplored. This study exhaustively evaluates all possible entanglement topologies of a TwoLocal feature map and benchmarks them against classical methods and standard quantum feature maps on two synthetic datasets (Adhoc3_150 and Adhoc4_150) and two real-world datasets (Blood and Banknote). On Adhoc4_150, the best TwoLocal topology achieves an accuracy of 0.7556, matching Pauli Z; on Blood, it reaches 0.6889, exceeding RBF SVM and Pauli Z at 0.6667; and on Banknote, it achieves 1.0000, compared with 0.9778 for RBF SVM. In contrast, Random Forest achieves the highest accuracy on Adhoc3_150 at 0.8222, exceeding all evaluated quantum configurations. Across all four datasets, increasing the number of entangling pairs does not consistently improve accuracy, while the standard linear, circular, and full Pauli ZZ topologies never outperform Pauli Z. To further characterize this behavior, we propose a factorial linear modeling approach that quantifies the contribution of individual entanglement edges and qubit connectivity to classification accuracy. On Adhoc4_150, two edges exhibit significant negative effects, whereas Blood and Banknote show both significant positive and negative edge effects. The q0q1 edge is significant across all three four-feature datasets but reverses direction across datasets. These results indicate that the effect of entanglement depends jointly on the feature-map architecture, dataset, and specific qubit connections rather than on entanglement density alone. The proposed framework can also be applied to sampled topology sets, providing a basis for targeted entanglement analysis beyond exhaustively enumerable low-qubit systems.

Beyond Reported Accuracy: A Verifiability-Focused, Multi-Axis Review of Content and Context Based Fake News Detection

Long Dinh Tuan — 2026-09-11

Fake news undermines information integrity, and the proliferation of large language models (LLMs) has intensified both its generation and its detection. Prior surveys catalogue methods but rarely verify the performance figures they synthesize, restrict comparison to same-benchmark settings, or assess risk of bias, so reported accuracies are often over-read. This verifiability-focused review analyzes an enumerated, non-exhaustive corpus of 45 content- and context-based detection studies (2018-2025; 42 quantitative-core) in which each reported headline metric is traced to its primary source and confirmed against it wherever the source is accessible (36 of 45; the nine paywalled studies are flagged as reported-but-unconfirmed), and released with the corpus. Studies are grouped into five method families, Transformer/PLM (31.0%), multimodal (26.2%), LLM-based (21.4%), graph/propagation (16.7%), and traditional machine learning (4.8%), and interpreted through a four-axis framework of evidence source, detection time, data dependence, and deployment objective. Without metric conversion or pooling, reported accuracies are strongly heterogeneous and not comparable across families (Transformer 74.8-99.36%, multimodal 82.4-99.97%, graph 84.4-94.36% or 0.75-0.93 AUC, LLM-based 48.6-89.0%); within Weibo, seven multimodal models span 82.4-91.8%, and GPT-4 attains 68.2% (not the ~95% often cited) on LIAR-binary. Three findings follow. First, reported performance appears shaped as much by data and evaluation design as by architecture, so architecture-only interpretations can be misleading; near-perfect results may reflect dataset-specific separability, leakage, or favorable evaluation designs rather than superior transferable capability. Second, the field has evolved cumulatively from static content cues toward multi-source evidence integration and reasoning, so the families are complementary evidence sources rather than exclusive choices. Third, no family is optimal across scenarios, and a central open problem is sustaining reliability under drift, new languages, attacks, and AI-generated content. We derive a dependency-ordered research agenda and, from a Vietnamese case study, transferable design principles for low-resource languages.

Explainable Machine Learning for Predicting Indonesian Vocational School Accreditation: Geographic Validation, Probability Calibration, and Subgroup Auditing

Muhamad Riyan Maulana — 2026-09-12

Vocational school accreditation is a high-stakes institutional classification, yet the geographic transferability and probability reliability of models based on routinely collected administrative data remain underexplored. This study developed a reproducible, leakage-controlled machine-learning framework for classifying the recorded accreditation classes C, B, and A among Indonesian vocational schools. Its main contribution is an integrated evaluation design that combines province-disjoint validation, probability calibration, cluster-aware uncertainty, and subgroup auditing to test reliability beyond conventional random-validation performance. A cross-sectional dataset containing 149,376 program-level records was deterministically aggregated into 14,391 schools, including 14,134 with valid labels. Eight provinces comprising 1,536 schools were locked before feature selection, model comparison, tuning, and calibration; 12,598 schools from 31 provinces supported development. The selected CatBoost model used 35 structural and vocational features with grouped out-of-fold temperature scaling. On the locked holdout, the calibrated model achieved balanced accuracy of 0.586, Macro F1 of 0.533, quadratic weighted kappa of 0.499, Macro ROC-AUC of 0.757, and expected calibration error of 0.049. School size, teacher resources, vocational staffing, program diversity, and status were most in-fluential. Class B had the lowest recall, and performance varied across school groups and provinces. The framework supports calibrated, auditable preliminary screening, but not automated accreditation or replacement of professional assessors.

Machine Learning Framework for SHM Alarm Log Triage under Alarm Flood and Limited Labeled Data: A Case Study on a Tropical Cable-Stayed Bridge

Aris Haris Rismayana — 2026-09-17

Structural Health Monitoring (SHM) systems on long-span bridges routinely generate alarm volumes exceeding operators' manual review capacity, a condition known as alarm flood, so that genuine structural anomalies risk being buried among far more numerous instrumentation faults. The problem is acute in developing-country settings, where monitoring falls to small operational teams rather than dedicated staff. This study proposes an alarm-log triage framework that classifies events directly from tabular database records and static sensor metadata, requiring no raw sensor time-series, offering an alternative to the raw-signal paradigm dominating bridge SHM. The dataset comprises 63,934 alarm records logged over 2.5 years on a 641.8-m tropical cable-stayed bridge: labeling collapses as volume rises, leaving only 401 usable labeled records from the pre-flood period, a weakly-supervised task under severe label scarcity. The pipeline integrates alarm-log attributes, causally-ordered temporal window features, and as-built sensor metadata, with fold-internal resampling and recall-oriented operating-point selection. Seven classifiers were benchmarked under identical stratified 5-fold cross-validation, including TabNet alongside linear, kernel-based, tree-ensemble, and neural baselines. Gradient-boosted tree ensembles separate clearly from all other families (F1-Macro 0.7991–0.8860 outside the tree family versus 0.9354–0.9432 within it; McNemar and DeLong tests confirm the separation, p < 0.05). Differences among the tree ensembles themselves are not consistently significant, so LightGBM is reported as the selected deployment model on consistent top-tier performance (F1-Macro 0.9432, AUC-ROC 0.9885 at the default threshold) rather than as demonstrably superior. A two-part ablation shows temporal alarm-rate features supply most of the discriminative power, while explicit imbalance handling yields no measurable benefit at the observed ratio, indicating that such mechanisms should be calibrated to imbalance severity rather than applied by default. The results establish operational alarm logs as a viable, lightweight input for structural event triage and document alarm flood quantitatively in a real bridge SHMS.

A Heterogeneous Multi-Task Deep Learning Approach for Geopolitical Risk Prediction and Cross-Country Heterogeneity in ASEAN Equity Markets

St. Dwiarso Utomo — 2026-09-23

Modeling the asymmetric transmission of geopolitical shocks across structurally diverse emerging markets presents a severe computational challenge. Traditional homogeneous deep learning approaches often suffer from negative transfer and base-rate bias when applied to cross-country financial prediction. To address this, we propose the Heterogeneous Multi-Task Spatiotemporal Network (HMT-STN), a structurally adaptive framework that integrates country-specific embeddings, heterogeneous task decoupling, and a Focal Loss mechanism. Evaluated on a daily panel dataset of ASEAN-4 equity markets (2016–2026), the HMT-STN is designed to simultaneously optimize directional return classification and volatility regression. Comprehensive evaluations using Precision, Recall, F1-score, and AUC-ROC demonstrate the framework's superior discriminative power, notably achieving 66.21% accuracy and an AUC-ROC of 0.68 in Thailand, a highly volatile transitional market. Rigorous ablation studies confirm that Focal Loss prevents catastrophic performance collapse in noisy environments, while country embeddings successfully isolate market-specific idiosyncrasies. Furthermore, SHAP (SHapley Additive exPlanations) analysis reveals a tripartite spectrum of cross-country heterogeneity, providing model-based feature attributions that are consistent with established economic theories of market segmentation and geopolitical risk sensitivity. Collectively, this study demonstrates that explicitly modeling cross-country heterogeneity through heterogeneous multi-task learning offers a robust, interpretable, and computationally efficient paradigm for geopolitical risk assessment in fragmented global economies.

Clinically Weighted Feature Fusion for Unsupervised Identification of Postpartum Depression and Anxiety Risk Subgroups Among Breastfeeding Mothers

Eka Prasetyaningrum — 2026-10-02

Postpartum depression and anxiety often go undetected, and breastfeeding experience is among the factors most strongly linked to both. However, previous research has largely been supervised, dependent on biomarkers, and rarely captures the heterogeneity of risk profiles among mothers with similar symptom levels. This study proposes Clinically Weighted Feature Fusion (CWFF), a clustering framework that weights clinical composite features based on their out-of-fold association with EPDS and GAD-7 scores, to map postpartum depression and anxiety risk subgroups among 2,010 breastfeeding mothers in the United Kingdom.

As a foundation, 22 composite features representing five clinical domains were compared with 30 raw variables using K-Means (k = 4). The engineered approach with equal weighting (equal-weight) showed better clustering quality, marked by a 28.2% increase in the Silhouette Score, a 21.8% decrease in the Davies–Bouldin Index, and a 26.8% increase in the Calinski–Harabasz Index; all significant based on bootstrap testing (p < 0.0001). Ablation analysis of leave-one-domain-out and leave-one-interaction-term-out (Friedman χ² = 794.15; p < 0.001) shows that this improvement mainly comes from Interaction Features and Mental Health History, not from all domains uniformly. This advantage is also only consistent for K-Means and Agglomerative Clustering, not for Gaussian Mixture Model (GMM) or HDBSCAN. Then, applying CWFF, the results show that CWFF raises Silhouette from 0.1537 to 0.2418 while also raising clinical validity (η²_EPDS from 0.071 to 0.096; η²_GAD-7 from 0.090 to 0.123), with all improvements significant (bootstrap 95% CI does not include zero). These findings indicate that clinically informed weighting improves risk-subgroup identification compared to equal-weight, potentially supporting low-cost postpartum screening in primary healthcare settings.

QE-CCHM: A Quantum-Walk-Enhanced Cross-Coupled Hyperchaotic Map with Broad Parameter-Domain Exploration and Its Application to Image Encryption

De Rosal Ignatius Moses Setiadi — 2026-10-18

Chaotic and hyperchaotic systems have been widely investigated for pseudorandom sequence generation and image encryption, but many existing approaches remain limited by relatively narrow parameter regions and predominantly classical dynamics. This study proposes a Quantum-Walk-Enhanced Cross-Coupled Hyperchaotic Map (QE-CCHM), a hybrid classical–quantum dynamical framework that introduces process innovation by integrating nonlinear cross-coupled classical states with two discrete-time quantum walks (DTQWs) through bidirectional measurement feedback. The quantum observables dynamically influence the classical states and the coin angles of the coupled DTQW subsystems, making quantum evolution an integral component of the dynamical process. The proposed system is evaluated using phase-space trajectories, bifurcation diagrams, Lyapunov exponents (LE), and broad parameter-domain exploration. In the investigated region 3.5≤ξ≤50, both leading LEs remain positive across all 100 evaluated parameter values, increasing from 6.2501 and 5.3408 at ξ=3.5 to 9.5432 and 8.6661 at ξ=50. A global exploration further investigates ξ from 10^^-2 to approximately 10^³⁰⁸, including 10,033 evaluations across the extended positive-parameter domain. The resulting sequences exhibit high normalized entropy and low lag-1 autocorrelation across widely separated parameter scales. For application validation, the generated sequences are applied to a reversible image-encryption scheme combining pixel permutation and modular substitution. Across six grayscale images, entropy exceeds 7.998 bits/pixel, adjacent-pixel correlations approach zero, and NPCR and UACI range from 99.59% to 99.63% and 33.09% to 33.69%, respectively, demonstrating the potential of QE-CCHM for pseudo-random sequence generation and image encryption.

NAA-GAN: Noise-Aware Cross-Attention for Robust Blind Image Watermarking

Surni Erniwati — 2026-10-25

Robust blind image watermarking remains challenging because embedded watermark information must remain recoverable after image processing while preserving the visual quality of the watermarked image. This paper proposes NAA-GAN, a generative adversarial watermarking framework that conditions the encoder on a six-dimensional descriptor of the anticipated distortion category. The encoder integrates dense feature aggregation, noise-aware cross-attention, and residual embedding to insert a 30-bit watermark. During training, a noise subnetwork simulates image processing attacks, while the decoder recovers the watermark from the received image without requiring the distortion descriptor. Experiments on COCO, PASCAL VOC2012, ImageNet, and EUVP assess image fidelity using peak signal-to-noise ratio (PSNR) and structural similarity index measure (SSIM), and robustness using bit accuracy (BA). NAA-GAN achieves PSNR values ranging from 36.96 to 37.79 dB and average BAs above 98% across the four datasets under the five evaluated attacks. On COCO, it achieves 99.02% BA under JPEG compression and an average BA of 92.59% across distortions not encountered during training. An ablation study shows that removing the distortion descriptor reduces COCO PSNR from 37.38 to 36.73 dB and average BA from 98.07% to 96.14%. Across all four datasets, NAA-GAN achieves higher PSNR and JPEG BA than the compared models. However, ARWGAN-CN achieves higher SSIM and average BA, indicating that the proposed method improves pixel-level fidelity and JPEG robustness but does not consistently outperform competing methods across all evaluation metrics.

A Five-Step Multi-Dimensional Learnability Validation Protocol for Meningitis Clinical Risk Assessment

Fathimah Al-Ma'shumah — 2026-10-25

Clinical machine learning is routinely judged on discrimination alone. A model that separates classes well is reported as successful, even when its probabilities are miscalibrated, its explanations point at variables with no measured association with the outcome, and nobody has checked whether the target can be learned from the available features at all. This study proposes a five-step multidimensional learnability validation protocol that tests a prediction target on five axes: feature-target dependence before modelling, calibration through Brier score decomposition, clinical utility through decision curve analysis, interpretability through SHapley Additive exPlanations, and agreement between fitted coefficient signs and the directions reported in the clinical literature. The protocol is the contribution; no new learning algorithm is proposed. Five established classifiers, Logistic Regression, Random Forest, LightGBM, CatBoost, and XGBoost, are used as instruments on a public synthetic meningitis dataset of 1,200 records with 10 clinical predictors and three prediction targets. Under repeated stratified cross-validation with five folds and five repeats, the four tree-based models reached mean areas under the curve of 0.969 to 0.971 and mean Matthews correlation coefficients of 0.780 to 0.793 for aetiological diagnosis, and 0.969 to 0.972 and 0.791 to 0.802 for clinical risk level; both targets passed all five validation steps. The mortality target failed all five: the strongest feature-target correlation was 0.067 and the strongest mutual information 0.013, the Brier score of the Platt-scaled reference model (0.0949) did not fall below the uncertainty baseline (0.0944), no threshold between 0.05 and 0.50 gave net benefit above both default strategies, the leading SHAP attribution fell on a variable with no measured association, and only two of five literature-referenced coefficient signs agreed. Judged only by whether mean discrimination exceeded chance, mortality could have appeared marginally viable. The protocol is demonstrated, not externally validated, and a target that passes is eligible for further validation rather than ready for clinical use.

Lightweight Hybrid Deep Learning Model for Efficient Edge-Cloud Computing: A Case Study on Poultry Disease Detection

Yudi Susanto — 2026-10-25

Poultry disease classification from dropping images remains challenging in resource-constrained farming environments, where visually similar symptoms complicate diagnosis and limited connectivity makes raw-image transmission costly. Although deep learning can automate classification, achieving a practical balance between model compactness, diagnostic performance, and communication efficiency remains a challenge. This study proposes a hybrid edge–cloud framework that combines a modified SqueezeNet backbone with Global Concat Pooling (GCP) for feature extraction and a Trio-Boosting classifier comprising XGBoost, LightGBM, and CatBoost with Weighted Soft Voting. Feature extraction is performed locally, while compact feature vectors are transmitted to a cloud endpoint for classification. Experiments show that GCP provides the principal performance improvement in the staged ablation, substantially outperforming conventional Global Average Pooling, whereas Trio-Boosting offers only a marginal gain over CatBoost alone. Comparisons with simpler classifiers likewise reveal no statistically significant advantage for the ensemble. In a separately trained final-model run, the framework achieved 89.23% accuracy and a weighted F1-score of 0.8924 on the test set; however, this result was not reproduced consistently during retraining and should therefore be interpreted cautiously. The backbone contains approximately 0.39 million parameters and demonstrates performance comparable to MobileNetV2 under the evaluated protocol, although it does not match the accuracy of fine-tuned ResNet50. The split-inference design reduced the transmitted payload by approximately 75%, while measured end-to-end latency remained substantially higher than model-side computational latency because of network communication. These findings suggest that the framework’s primary value lies in compact feature extraction and reduced transmission payload rather than superior classification accuracy. Duplicate images and conflicting labels in the dataset, together with the use of a laptop-class CPU to emulate edge processing, limit the strength of generalization and deployment claims. Further evaluation on independent datasets and resource-constrained hardware under field conditions is required.

A Training-Free Symbolic Baseline for Interactive Medical Diagnosis: Automatic Rule-Slot Derivation and Weighted Evidence Combination

Naufal Denta Elsatya Nugraha — 2026-10-25

Interactive medical diagnosis requires sequential symptom elicitation before a diagnosis can be established. Although recent approaches employ reinforcement learning and transformer-based agents, their performance has not been adequately contextualized against a classical probabilistic baseline. This study introduces a training-free symbolic baseline that derives conditional symptom probabilities and three rule slots—must-have, exclusion, and critical—from co-occurrence counts in simulated patient records. Candidate diseases are ranked using observed evidence, while subsequent questions are selected by expected information gain. Evaluation on five benchmarks from a published reinforcement learning study reveals distinct performance regimes. On MedlinePlus (893 diseases), the proposed method achieves 99.27 ± 0.09% accuracy with an average of 1.75 questions, outperforming the published baseline (96.2 ± 0.1%; 3.41 questions). Across four SymCat variants, it improves accuracy by 3.03–7.70 percentage points but requires approximately twice as many questions; at a matched budget of 8.98 questions, the reference method achieves higher accuracy, highlighting a trade-off between diagnostic accuracy and interaction cost. Ablation experiments across six datasets show that removing weighted evidence combination reduces accuracy by 35.4–65.4 percentage points, while eliminating symptom inquiry reduces it by 23.3–66.8 percentage points. The safety veto removes up to 50.2% of candidates on average but changes the final diagnosis in at most 0.17% of consultations, indicating limited incremental impact on predictive accuracy under the evaluated configurations. Rule-base derivation takes 1.8 s for 893 diseases on a CPU. As all benchmarks rely on simulated patients, these findings establish benchmark-level performance rather than clinical effectiveness.

Conditional Returns to Language Adaptation: An Adaptation Ladder Study of Indonesian Short Answer Grading

Imam Rangga Bakti — 2026-11-01

Indonesian has several adapted language models for automatic short answer grading. Studies disagree about whether specialization is worthwhile; each test one point on the adaptation spectrum, so a losing language-specific model is indistinguishable from useless specialization. An ordinal comparison replaces it: a three-rung ladder of released checkpoints, from no Indonesian adaptation through regional to national continued pretraining, in two families whose rungs share a base model and parameter count. In the Gemma family, the first two rungs are siblings rather than ancestor and descendant, which the analysis carries through. Across 30,960 grading calls on 36 independent questions, with the question as the unit of inference, returns are conditional rather than cumulative. Regional adaptation helps both families, by 0.029 and 0.018 pooled quadratic weighted kappa (QWK). The second step adds little on the weaker base, 0.016, and harms the stronger one, −0.029. That decline is established in one family and directional in the other. Labelled examples drawn from the question being graded halve the remaining advantage, and examples from other questions attenuate it further, from +0.069 to +0.005, a difference of −0.064 [−0.107, −0.022], although they grade less accurately in absolute terms. Spreading the demonstrated score levels reduces it to +0.029. All six models moved in the direction predicted from their initial scoring bias at four shots, and five of six at eight shots. Error decomposition locates the gain in a shift in grading severity, not sharper discrimination, and the shift runs in opposite directions in the two families. A 124-million-parameter Indonesian encoder, included as a deployment reference, reaches 0.825 QWK in the social science domain at 596 MB, within 0.027 of the best 9-billion-parameter configurations. Language adaptation is therefore a conditional recalibration. Its value depends on the base model, on labeled answers already held, and on which score levels they display.

Benchmark-Defined Adaptive Candidate Reranking for Indonesian Multi-Hop Question Answering

Iqbal Abdul Rahman — 2026-11-01

Multi-hop question answering (MHQA) requires recovering a complete set of supporting documents, yet most retrieval-augmented pipelines apply one evidence budget and one selection strategy to every question. This paper studies benchmark-defined adaptive candidate reranking for Indonesian MHQA in a closed setting in which each question is paired with ten candidate documents. A routing classifier assigns each question to a two-gold-document branch (HotpotQA), processed by two-stage sequential reranking, or to a four-gold-document branch (a bridge-comparison subset of 2WikiMultihopQA), processed by parallel ranking-based reranking, using lightweight XLM-RoBERTa cross-encoders. All components are trained without stemming and selected on a held-out validation split. On 8,000 Indonesian-translated evaluation questions, the framework recovers the complete supporting-document set for 95.40% of questions with 3.00 documents per question and 97.69% document precision. In evidence recovery, it significantly outperforms a single-model fixed-K=4 reranker (+2.60 points; McNemar's exact test p < 10⁻¹⁷); relative to top-4 selection with the same branch-specific scorers, it trades 0.89 points of containment for 23.55 points higher document precision and one document fewer per question. Ablations attribute the largest effect to sequential conditioning (+14.00 points at an equal budget of two documents) and show that routing errors cost 0.03 points within the benchmark, that transfer initialization mainly accelerates convergence, and that Sastrawi stemming degrades the scorers. Class-wise probes and a cross-source test show that the routing classifier relies on a film-topic shortcut and misroutes 10.78% of unseen two-document questions; it is therefore a benchmark-defined controller, not a detector of reasoning complexity, and the results do not establish generalization beyond the evaluated benchmark. With GPT-5-mini as the generator, adaptive evidence uses 68.5% fewer context tokens than all ten candidates and 26.8% fewer than fixed-K=4 selection at 9 to 20 ms of additional selection latency per question, but no improvement in answer accuracy was found: answer accuracy remains statistically indistinguishable across evidence contexts and within the run-to-run variance of the generator.

Leakage-Aware Cross-Dataset Multi-Task Learning for Shrimp Disease Recognition and Biomass Estimation

Budi Juarto — 2026-11-09

Public shrimp image collections are reused to train disease classifiers and non-invasive biomass estimators, yet they mix originals with published augmentations, byte-identical copies, and repeated views of the same animal, potentially inflating performance through data leakage. This study examines what these collections support once leakage is controlled: how protocol and acquisition conditions, rather than architecture alone, influence performance; whether disease recognition transfers to an independently collected dataset; and whether monocular imagery supports individual-weight estimation. One encoder is trained across three public datasets under masked supervision, using a grouping strategy that accounts for duplicates, augmentations, near-duplicates, and repeated views; a 22-check audit guides training and evaluation, with 240 runs reported. Regrouping by animal rather than file eliminates the 84.8–92.6% animal overlap in ShrimpDiseaseImageBD test folds, reducing it to 0.0%. A random image-level split yields 0.976 balanced accuracy, but 63% of its test records are published augmented copies; using originals alone yields 0.947, compared with 0.934 in the augmentation-enabled group-aware arm, which tests different original records. Transfer is asymmetric: the best-performing encoder reaches 0.615 balanced accuracy and 0.751 area under the curve on TigerShrimpBD; on ShrimpDiseaseImageBD, every encoder ranks below chance, with values of 0.271–0.390. TigerShrimpBD's 0.998 within-dataset score is affected by acquisition confounding, as every recovered capture session contains only one health class. Holding out a physical site increases biomass error from 3.18 g to 9.20 g, indicating limited cross-site generalization. The contribution is a reusable audit and evaluation protocol for interpreting performance on such collections.

An Interpretable Neuro-Fuzzy Framework for Mental Health Indicator Classification with Contextual Representation and Metaheuristic Optimization

Haeni Budiati — 2026-11-10

Social media text provides valuable information for analyzing mental health-related expressions. However, contextual language variation, class imbalance, and limited model interpretability remain important challenges in text classification. This study evaluates different text representations, class-imbalance handling strategies, and metaheuristic optimization within an Adaptive Neuro-Fuzzy Inference System (ANFIS) framework. A total of 19,429 Indonesian-language posts were retained after duplicate removal and label-conflict adjudication. The dataset was partitioned using a fixed group-aware split into 13,600 training, 2,914 validation, and 2,915 test records, covering four categories: Distress, Crisis, Protective, and Situational. The experiments examined lexical-statistical, distributed-semantic, and contextual representations using Plain ANFIS, class-balanced focal ANFIS (CBF-ANFIS), and its metaheuristic-optimized variants based on the Whale Optimization Algorithm (WOA) and Crow Search Algorithm (CSA). Model performance was evaluated across three random seeds and compared with a TF-IDF-based Logistic Regression baseline and direct IndoBERT fine-tuning. Direct IndoBERT achieved the highest test Macro-F1 of 0.7685 ± 0.0046, followed by preliminary Plain ANFIS (0.6942 ± 0.0011) and final CBF-ANFIS (0.6897 ± 0.0054). Although WOA and CSA improved validation Macro-F1, their mean test Macro-F1 scores were slightly lower than that of CBF-ANFIS (0.6890 and 0.6859 versus 0.6897, respectively), indicating no consistent generalization benefit. Overall, direct IndoBERT provided superior classification performance, whereas the neuro-fuzzy framework offered a structured mechanism for tracing rule activation and inference in the latent feature space. Nevertheless, these rules do not provide direct linguistic or clinical interpretations.

When Similar Models Disagree: Model Multiplicity and Geographic Robustness in School Meal Utilization Prediction

Amanatullah Pandu Zenklinov — 2026-11-10

Competing predictive models can achieve similar aggregate performance while producing different classifications and downstream conclusions for the same individuals. This model disagreement raises an important question about whether conventional performance metrics adequately capture the reliability of model-based analyses. This study investigates model-choice disagreement across multiple analytical layers using voluntary-response survey data from 8,802 respondents who reported receiving meals through Indonesia’s Free Nutritious Meals program. Logistic Regression (LR), Random Forest, and XGBoost were evaluated using a nested school-disjoint validation framework. All three models were compared in terms of discrimination and probability calibration, while the competitive models were further examined through high-risk classification, predictive attribution, model-based scenario analysis, and province-holdout robustness. LR and XGBoost achieved nearly identical PR-AUC values (0.7270 and 0.7277, respectively), with similarly close pooled PR-AUC under province holdout (0.7288 and 0.7285). Despite this aggregate similarity, the models assigned different high-risk classifications to 5.41% of respondents. Their scenario analyses also produced different threshold-crossing rates: 8.87% of eligible respondents under LR (382 of 4,307) and 6.66% under XGBoost. Random Forest exhibited weaker probability calibration in the fitted configuration and was not retained for downstream comparison. Overall, the findings reveal that model agreement at the population level does not necessarily extend to individual classifications and scenario outputs. LR was selected as the reference model through documented judgment considering predictive performance, calibration, robustness, parsimony, and interpretability. These results demonstrate the importance of evaluating model-choice robustness beyond conventional predictive performance metrics, particularly when model outputs inform downstream interpretation or classification.