Text-to-Visualization (Text-to-Vis) systems span three paradigms–rule-based, neural, and Large Language Model (LLM)-based–each emitting incompatible output artifacts, from declarative grammars to imperative plotting code, so existing benchmarks compare systems only within a single output family, and unified Abstract Syntax Tree (AST) extraction introduces systematic measurement artifacts mistaken for capability gaps. We propose the Multi-Level Format-Agnostic (MLFA) evaluation framework, which assesses systems at three deliberately disaggregated levels: executability through rendering (L1), semantic accuracy via LLM-based component extraction calibrated against ground-truth parsers (L2), and visual fidelity via multimodal-LLM comparison of rendered images (L3). Validated on seven systems spanning all three paradigms across two benchmarks, L2 recovers capability that format-bound axis metrics had hidden in code-emitting LLM systems, L3 yields a uniform visual-fidelity ranking across paradigms, and an efficiency analysis exposes token and latency gaps favoring specialised neural architectures. To our knowledge, MLFA is the first evaluation framework to apply multi-level measurement uniformly across heterogeneous output formats, establishing disaggregated measurement as a prerequisite for fair cross-paradigm comparison.
An autoencoder (AE) is a neural network that reconstructs its input. AE consists of two sub-networks: the encoder and the decoder. AE has been commonly applied in feature extraction and anomaly detection. However, AE was not applied to AE models themselves. Accordingly, this paper develops AE for AEs, namely AEAE. For this purpose, we apply AE to the parameters (weights and biases) of AEs. The experiment includes 1) computing reconstruction error of the AEAE model, 2) applying feature extraction and classification to AE models, and 3) one-class classification of AE models using the reconstruction error of the AEAE model. The most interesting outcome is that biases of output layers are distinguishable with high accuracy, whereas other parameters are hard to classify. Moreover, AEAE could be extended to wider directions, applying deep learning to deep learning models. Finally, we got another question: how to apply AE to AEAE models.
The rapid growth of short-video platforms has created a content environment where information spreads quickly and user engagement changes greatly over time. As a result, early trend detection has become an important task for intelligent software analytics systems. In this study, we propose an analytical framework for detecting trends in short videos on social media using multi-source features and machine learning models. We construct a dataset of 1,922 TikTok videos, of which 172 are labeled as trending based on engagement growth criteria. For each video, we extract features related to metadata, content creators, posting time, captions, background music metadata, contextual information, and semantic representations of captions. Several machine learning models are evaluated, including Logistic Regression, Random Forest, SVM, LightGBM, and XGBoost. The results show that XGBoost achieves the best performance, with an F1-score of 0.8367 for the Trending class and a Macro-F1 score of 0.9108. The results also show that F1-score and PR-AUC are more informative than accuracy alone for this imbalanced trend detection task. Overall, this study shows that combining multi-source features with boosting models is a feasible approach for short-video trend detection. The dataset and source code are available upon reasonable request for research purposes.
Effective coordination between interacting actors relies on messages whose reliability varies with operational conditions, strategic incentives, and previous interactions. Assessing a message from its current content alone may therefore be insufficient, particularly when actors adapt their communication over time. This paper proposes a credibility-aware framework that formulates strategic information sharing as a sequential prediction problem, combining current messages, operational variables, observed actions, and interaction histories to estimate the reliability of communicated information and support coordination decisions. A neural sequential learning model learns a dynamic representation of actor credibility, capturing temporal dependencies and changes in communication strategies. Compared with Bayesian updating and memoryless prediction across simulated scenarios, the approach is evaluated on both predictive performance and the consequences of predictions for decision quality and coordination outcomes. The study investigates whether data-driven sequential prediction can provide adaptive, robust credibility estimates when communication patterns are uncertain, context-dependent, or time-varying.
As a by-product of the rapid advancement of big data and artificial intelligence, high-dimensional data has been expeditiously evolving in both quantity and complexity, resulting in the critical need for suitable, effective retrieval and management systems. With this motivation, numerous classical solutions to the problem of high-dimensional data search were proposed. One of the most well-known instances of these is locality-sensitive hashing (LSH) and its prominent ability to run in sub-linear time. In situations with large databases, candidate validation, the last step of LSH’s procedure, can present itself as a bottleneck, which, in most cases, induces a tradeoff between accuracy and efficiency. Therefore, in this paper, we intend to utilize the renowned ability to speed up specific tasks like search problems of quantum computing to alleviate this issue and develop a quantum locality-sensitive hashing (QLSH) system. Specifically, by exploiting a Hamming-distance-calculating quantum circuit to perform candidate validation, we reduce the complexity of this step while preserving the retrieval performance compared to the classical counterpart in approximate search among high-dimensional data, as demonstrated by quantum simulation results.
Urban trees play a key role in carbon sequestration, but estimating their carbon stock at fine spatial scales is challenging in complex urban environments. This study develops a web-based system to estimate and visualise urban tree carbon stocks using satellite imagery, machine learning, and geographic information systems (GIS). Sentinel-2 multispectral data and tree-level field survey measurements were used as input. Ground-based carbon estimates were derived using allometric equations and linked with spectral features extracted from satellite imagery. Several machine learning models, including Random Forest and CatBoost, were evaluated using spatial cross-validation to reduce spatial bias, and an ensemble averaging model was selected based on overall performance. The system supports both single Area of Interest (AOI) analysis and AOI comparison for change detection, allowing users to define AOIs interactively and visualize results through spatial dashboards. Model evaluation indicates stable performance for low to medium carbon trees, with increased uncertainty for larger trees due to known limitations of optical remote sensing. In general, the system demonstrates the practical integration of remote sensing, machine learning, and GIS for urban carbon assessment and supports applications in urban planning and environmental management.
This paper presents a dual-stream Transformer-based Visual Question Answering (VQA) framework for brain tumor MRI interpretation, paired with a custom 4,000-pair clinical question-answer dataset generated with LLM assistance and manually verified for medical plausibility. Two configurations, BERT+ViT-B/16 and RoBERTa+BEiT, were trained under identical conditions on 1,000 MRI images (250 per class) spanning glioma, meningioma, pituitary tumor, and no-tumor categories. On an 800-pair held-out test set, BERT+ViT achieved 82.75% accuracy (WUPS 0.8281, F1 0.1007) and RoBERTa+BEiT achieved 82.38% (WUPS 0.826, F1 0.1047), both reaching 100% on binary questions and approximately 66% on open-ended questions. Error and prediction-distribution analyses show both the flawless binary scores and the persistent open-ended errors are largely artifacts of the dataset’s templated answer structure, meaning much of the headline accuracy reflects dataset regularities rather than genuine visual-linguistic reasoning. We discuss this bias and propose radiologist involvement in future QA-pair design as the primary route to a less biased, more clinically meaningful benchmark.
This study presents the MASEC-Feedback Model, a multi-agent diagnostic framework designed to simulate workplace interaction and identify functional communication gaps. Interaction logs and diagnostic reports were collected from 149 international practitioners from Vietnam, Myanmar, Sri Lanka, and Nepal, with proficiency levels ranging from A2 to B2. The study provides theoretical insights into Knowledge Engineering for Precision Education. It requires a paradigm shift from reactive error correction to proactive, system-level diagnosis of professional readiness. The analysis identifies a separated two-layer architecture that bridges pedagogical disconnect, the localization of a high-volatility B1 performance breakpoint, and the reduction of algorithmic agreeability bias through weighted multi-agent consensus. These findings provide a strategic roadmap for deploying intelligent multi-agent scaffolds to optimize the unique developmental pathway of professional practitioners in the digital age.
Unexpected equipment failure in heavy diesel engines leads to prohibitive operational costs and critical downtime during essential seasonal windows. This paper presents an integrated predictive maintenance (PdM) framework that combines a 6-state Finite State Machine for real-time monitoring with a multi-model machine learning ensemble. The architecture utilizes Random Forest for fault classification, Isolation Forest for vibration anomaly detection, and Linear Regression for oil pressure forecasting. Key design innovations include a temporal debouncing mechanism to filter transient noise, standardized Diagnostic Trouble Codes for technician interpretation, and a priority-based maintenance scheduling algorithm that weighs failure severity and asset criticality against estimated remaining useful life. The proposed framework was evaluated using three public benchmark datasets: Scania APS, NASA C-MAPSS, and CWRU Bearing. The system achieved accuracies of 99.3% on Scania APS, 96.2% on NASA C-MAPSS, and 98.4% on CWRU Bearing. On the Scania APS dataset, which serves as the closest operational proxy to agricultural diesel hardware, the model maintained a 0.2% false positive rate. This research contributes a systems-integration methodology demonstrating how heterogeneous AI techniques can be synergized for specialized machinery. While validated using public benchmarks, future work will focus on deploying this framework using live telemetry from engine hardware in real-world environments.
The integrated rail transit system in Kuala Lumpur, especially such as the Light Rail Transit (LRT), Mass Rapid Transit (MRT), Keretapi Tanah Melayu (KTM), Monorail, and Express Rail Link (ERL), plays an important role in supporting urban connectivity by connecting commercial centers, residential areas, and major destinations. However, existing transit applications provide a fragmented user experience, static functionality, and poor interface design. They currently lack a one-stop app experience with features to analyse ridership trends and discover nearby attractions. To address these issues, this project presents KitaRide, a web-based application that integrates spatio-temporal visualisation and machine learning techniques to analyse ridership patterns and the nearest attractions in Kuala Lumpur. A random forest regression model was utilised to predict ridership patterns based on temporal and contextual features. It returned 12% for MAE, 2% for MSE, and 14% for RMSE. A Power BI dashboard was also embedded into the application to visualise trends using line charts and heatmaps. The system evaluation involved black-box testing, where all test cases passed successfully, and a System Usability Scale (SUS) score of 80.83, indicating high usability and user acceptance. Despite its effectiveness, the system is limited by the absence of station-level ridership data, real-time trend patterns, and live service disruption reporting. Finally, future enhancements may include expanded datasets, real-time data integration, and an additional feature for user-reported station status.
High-Performance Computing (HPC) clusters are essential infrastructure for scientific computing and AI workloads. As they grow larger and more heterogeneous, effective monitoring becomes necessary to diagnose bottlenecks and understand resource usage. Existing HPC monitoring tools, however, expose telemetry mainly through static dashboards or hand-written SQL, Flux and PromQL queries. Such interfaces serve predefined questions but make ad-hoc operational reasoning difficult, particularly when operators must identify which node, process, job or user is responsible for abnormal behavior. This paper proposes VizMon, an HPC monitoring system with an integrated natural-language visualization agent. VizMon connects natural-language questions to its modular monitoring substrate, so that non-expert users can interrogate the system directly. Guided by the temporal scope of each question, the agent dispatches it to either a historical pipeline that returns a static analytical chart or a real-time pipeline that returns a live Grafana panel. Experiments show that the substrate stays under 2% of a 32 GiB host across four measured cost drivers, and that on the public VisEval benchmark VizMon’s historical chart-generation core reaches a 78.3 % pass rate and a 3.46 composite quality score, outperforming four reference systems at moderate token and latency cost.
In-context learning (ICL) is a fundamental capability of large language models (LLMs), yet its reliability remains difficult to assess before deployment, especially for locally deployed open-source models. Standard evaluation often focuses on task accuracy, but ground truth-based evaluation alone cannot capture whether model behavior remains consistent under controlled input changes. Metamorphic testing (MT) addresses this limitation by checking whether predictions satisfy predefined metamorphic relations under input perturbations, even when exhaustive ground-truth labels are unavailable. While prior work such as MMT4NL has shown the promise of this approach, existing MT studies have mainly focused on commercial LLMs. This paper fills this gap by extending the MMT4NL framework to five open-source LLMs served through Ollama using three perturbation categories: fairness-related demographic perturbations, semantic-equivalence negation, and semantic-inequivalence negation. The last category is introduced to test whether models appropriately change their predicted sentiment when a negation transformation changes the meaning of the input. Results show that open-source models exhibit different consistency patterns across perturbation types: they are relatively stable under fairness-related perturbations involving gender and geography, but show weaker consistency under negation-based transformations, especially semantic-inequivalence negation. These findings suggest that metamorphic testing can provide a useful pre-deployment diagnostic tool for identifying behavioral inconsistency in open-source LLMs.
Formalization of informal use cases to series of structured functional requirements (FRs) is an important step in developing dependable software. However, such conversion process is commonly labor-intensive and slow. In this work we present a fully automated pipeline that translates use-case flowcharts (as SVG and images) into rule based FR documents via a five-step process: input validation, OCR/SVG text extraction, structure detection, local large language model (LLM) drafting using Ollama, and report production. The manual use-case-to-FR workflow is compared with our automated pipeline, revealing the points in which ambiguity and omissions are likely to occur. We prepare a relatively smaller data set composed of use cases and attempt to generate structured functional requirements. To achieve this, we tested our proposed pipeline on a set of two experimental settings: one with a single SVG-based user flow to establish basic behavior and another with six SVG use-case flows at once for providing an initial assessment of robustness and scalability. This two-step evaluation facilitates a systematic analysis on the individual components’ performance as well as overall effectiveness with growing input complexity. Preliminary evaluation demonstrates on two independent reports: an overall 80.5% accuracy for a one-flowchart batch; and an overall 87.0% accuracy for six-flowchart batch. On both batches structure detection reaches 100.0% performance, while text extraction is still the main bottleneck (between 50.0% and 66.67%). The pipeline produces traceable normalized FR statements and a PDF (with issues per file and suggestions) for an analyst to review and fine-tune.
Wildfire monitoring using Uncrewed Aerial Vehicles (UAVs) requires accurate pixel-wise segmentation under difficult conditions (including dense smoke, fragmented flame structures, and substantial scale variation). In this paper, we present Flame Context Routing–Fusion Network (FCRF-Net), a context-aware segmentation model for UAV wildfire detection. The proposed approach employs Adaptive Context Routing Module (ACRM) to refine encoder features using cross-scale Feature-wise Linear Modulation, parallel local–global feature extraction, and uncertainty-guided spatial routing. It maintains fine-boundary details in confident regions while emphasizing global context in ambiguous areas. Afterward, we use a Gated Dual-Fusion Refinement (GDFR) decoder that can progressively restore spatial resolution by adaptively fusing local and dilated features, aided by auxiliary supervision for stable training and improved detail reconstruction. Experimental results on two datasets, FLAME and Boreal Forest Fire, show that FCRF-Net reliably outperforms previous baselines. On the FLAME dataset, it achieves an IoU of 0.8352, a precision of 0.9068, a recall of 0.9135, and an F1-score of 0.9102. On the Boreal Forest Fire dataset, our proposed technique can achieve an IoU of 0.8607, a precision of 0.9298, a recall of 0.9206, and an F1-score of 0.9251. Ablation studies show that ACRM and GDFR provide complementary gains in the segmentation accuracy.
This paper presents a lightweight Vision–Language Model (VLM) designed to support real-time indoor healthcare applications under resource-constrained deployment conditions. The proposed architecture integrates a ResNet-18 visual encoder and a BERT-base (uncased) text encoder to learn a compact multimodal representation for binary image–text matching, where a pair is classified as match or mismatch. A dataset of 600 image–caption pairs was constructed from open-source web repositories using four caption templates describing common indoor healthcare-like scenes and spatial relations. Positive and negative pairings were generated and split into 480 training and 120 validation samples. To assess generalization beyond the training distribution, an external test set was created using RGB images captured from a depth camera, paired with slightly rephrased prompts to evaluate robustness to linguistic variation. Experimental results on the test set demonstrate an overall accuracy of 84%, with a precision of 95%, a recall of 72%, and an F1-score of 82%, indicating reliable matching performance. In addition, an explanation-oriented trustworthiness evaluation was conducted using two questions, showing stronger performance for spatial-relation reasoning than for explicit human presence verification, reflecting the model’s scene-level matching formulation rather than object-level grounding. Overall, the proposed lightweight VLM provides an efficient baseline for context-aware indoor healthcare monitoring and motivates future work on integrating lightweight grounding modules to improve object-centric reliability in safety-relevant surveillance scenarios.
Performance regression testing in industrial environments often involves large and costly test suites, making full execution impractical for every regression cycle. This challenge is more pronounced in black-box settings, where only test metadata and historical failure information are available. This paper presents an unsupervised, cluster-based approach for test subset selection in black-box performance regression testing. The method groups tests using their metadata similarity and selects representative tests based on historical failure coverage. Evaluation on an industrial regression suite of approximately 3,900 tests demonstrates a clear trade-off between failure detection capability and execution efficiency. The results show that the proposed approach significantly reduces execution time while preserving a substantial portion of failure coverage and maintaining stable performance on unseen regression cycles. These findings indicate that the approach is suitable for practical deployment in continuous integration (CI) environments.
Accurate legal question answering demands navigating hierarchical document structures, following cross-references, and multi-hop reasoning, capabilities that standard Retrieval-Augmented Generation (RAG) systems lack. Vietnamese legal text compounds this challenge through tonal ambiguity, domain-specific abbreviations, and rigid structural hierarchies. This paper presents a three-tier knowledge graph RAG architecture that integrates a domain ontology, a knowledge graph, and a document tree forest. Three retrieval tiers, LLM-guided tree traversal, dual-level semantic search with intent-aware Personalized PageRank, and Reciprocal Rank Fusion with knowledge graph expansion, operate over these representations. Evaluation on 530 expert-annotated questions across enterprise law, traffic safety law, and postgraduate education regulation shows that the proposed system retrieves at least one correct article in the top 10 results for 92.2% of queries, outperforming the strongest baseline by 7.4%. These results hold across all three domains without domain-specific tuning.
Scaling up detection models is commonly assumed to improve accuracy. However, this study shows that larger models may reduce sensitivity to small defects in printed circuit board inspection. A systematic evaluation of five models on HRIPCB reveals that larger YOLOv8 variants improve overall localisation accuracy, measured by mAP50-95 (0.643→0.761), but they do not consistently improve APsmall, the most critical metric for micro-scale PCB defects. Compact YOLOv8n (0.587) achieves better small-object detection than YOLOv8m (0.563), YOLOv8l (0.552), and YOLOv8x (0.556). To address this trade-off without sacrificing speed, this study proposes a Knowledge Distillation (KD) framework that transfers temperature-scaled classification logits and multi-scale neck features from a YOLOv8m teacher to a YOLOv8n student. Combined logit and feature distillation with P3-scale-weighted feature alignment achieves the best result (mAP50-95 = 0.662, APsmall = 0.617). KD shows larger relative gains in the from-scratch regime (+1.8%) than with COCO pretraining (+1.4%), confirming that KD is most valuable when the student lacks prior representations. This ablation study identifies when and why KD provides modest but consistent improvements in compact PCB inspection models.
Dialogue-based Conversational Recommender Systems (CRS) are trained to reproduce item sequences chosen by human annotators, who disproportionately mention popular titles. The model learns to mimic this bias, leaving most of the catalog unrecommended, a problem known as low catalog coverage. We propose Soft-Rank Diversity (SRD), a fully differentiable training loss that penalises concentration of probability mass on clusters of similar items, encouraging the model to spread recommendations across the catalog. We target knowledge-graph-based CRS because knowledge graph (KG) embeddings place items that share real-world attributes close together, so SRD can detect and penalise real redundancy among candidates. SRD attaches to the item-scoring step of multi-hop KG reasoning, where it operates on a focused subset of related candidates, and admits an efficient O(MdB) factorisation, where M is the number of candidate items, d is the embedding dimension, and B is the batch size. The closest alternative training loss, the Determinantal Point Process (DPP) likelihood, incurs cubic cost and, as we show, degrades both accuracy and coverage. On CR-Walker with the ReDial and TG-ReDial benchmarks, SRD improves catalog coverage with competitive recall on ReDial and an accepted accuracy-coverage trade-off on TG-ReDial, while requiring nearly half of training time compared to DPP.
Automated diagnosis of tuberculosis and pneumonia from chest X-ray images is a challenging problem since the two diseases often present similar or indistinct imaging. In this study, a hybrid deep learning model was proposed by combining ResNet50V2 with left-right multi-head attention (LRMH) model. The main idea is to take advantage of the natural lateral structure of the lungs to better focus on clinically significant abnormal areas. For a fair comparison, the proposed model was compared with several different baselines including DenseNet121, EfficientNetB0, ResNet50V2, ResNet50V2 combined with generic attention and ConvNeXtTiny. Experimental results on publicly available chest X-ray datasets showed that the proposed model achieved the best overall results with an accuracy of 94.5% and F1 score of 94.7%. More importantly, the model in the tuberculosis class had a recall rate of 95.0% and an F1-score of 94.9% which was higher than the other comparison models. These results indicate that the LRMH mechanism can be beneficial in reducing missed tuberculosis cases, and in improving the reliability of automated chest x-ray screening systems.
