Abstract: Urban planners need continuous, scalable methods to evaluate pedestrian Level of Service (LOS). Static and locally calibrated approaches fail to capture the dynamic, network-wide, and context-dependent nature of pedestrian activity. While traditional LOS uses fixed density thresholds and data-driven models predict continuous flows, neither supports cross-city analysis due to context-specific assumptions. This study introduces a transferable analytical framework for predicting pedestrian LOS using large scale urban sensor data that captures both recurrent temporal demand patterns and spatial dependencies within street networks. The framework is evaluated using pedestrian sensor data from three cities Melbourne, Dublin, and Zurich, which represent diverse geometries, demand profiles, and sensing infrastructures. Results show strong in-domain Melbourne performance (accuracy 79.7%; Acc±1 99.1%) and effective cross-city generalization. Few-shot fine-tuning with only 5% labeled target-city data recovers 95–99% of in-domain performance, demonstrating practical scalability. KernelSHAP explainability reveals short-term temporal lag features universally dominate predictions, while spatial/contextual factors exhibit city-specific influence tied to local morphology. These findings demonstrate transferable GeoAI methods can support real-time pedestrian congestion monitoring and evidence-based public-space management, offering planners a scalable decision-support tool to enhance walkability, safety, and equitable access to high-quality public spaces in contemporary cities.

Abstract: Accurate vessel speed prediction is essential for maritime traffic supervision, navigational safety, and intelligent coastal management. However, due to the nonlinear, time-varying, and context-dependent characteristics of vessel motion in nearshore waters, conventional single-model approaches often fail to provide sufficiently accurate forecasts. To address this issue, this study proposes a hybrid deep learning framework for AIS-based nearshore vessel speed prediction and risk warning, integrating a temporal convolutional network (TCN), an attention mechanism, and a bidirectional long short-term memory network (BiLSTM) into a unified architecture. In the proposed framework, TCN is used to extract local temporal patterns and multi-scale sequence features from historical AIS observations, the attention mechanism is introduced to adaptively emphasize informative representations, and BiLSTM is employed to model bidirectional contextual dependencies in vessel motion sequences. On this basis, a speed-risk warning process is constructed by combining the predicted speed with electronic-fence threshold constraints. Experiments conducted on real AIS data from coastal waters show that the proposed method outperforms several benchmark models in terms of mean absolute error (MAE), mean squared error (MSE), root mean square error (RMSE), and coefficient of determination (R2). The results demonstrate that the proposed framework can effectively improve vessel speed prediction accuracy and provide practical support for proactive maritime supervision and nearshore safety management.

Abstract: We deploy an intrusion detection classifier on the STM32N6570-DK, a Cortex-M55 MCU with the Neural-ART NPU. Using the approximate T = 1 SNN–INT8 ANN equivalence, we compile a lightweight MLP to the NPU and evaluate four datasets: NSL-KDD (5-class), UNSW-NB15 (10-class), CICIDS2017 (15-class), and IoT-23 (5-class). Results are reported as mean ± std over multi-seed runs (5–20 seeds), with paired Wilcoxon signed-rank tests and Holm–Bonferroni correction. Across all datasets, INT8 NPU inference runs in 0.29–0.46 ms (2.7–4.2× faster than the same model on Cortex-M55 CPU), with estimated energy 44–69 μJ per inference and Flash 105–138 KB. Compared with recent MCU-class deployments on STM32F7 (31 ms, 7.86 mJ) and Raspberry Pi 3B+ (27 ms), our path delivers 59–107× lower latency; the estimated energy envelope implies 114–179× lower energy than STM32F7. QCFS and ReLU are statistically indistinguishable on all four datasets (p ≥ 0.227), supporting practical T = 1 near-equivalence under commodity MCU deployment constraints. Energy is estimated from STMicroelectronics application note AN5946 rather than direct on-board measurement, and UNSW-NB15 shows greater INT8 quantization fragility than NSL-KDD. We frame this as a deployment case study on a commodity Cortex-M-class MCU paired with a general-purpose NPU (Neural-ART), bounded by a documented systematic literature search (Supplementary File S1).

Abstract: The systematic construction of expansive fictional universes, known as worldbuilding, faces significant challenges in maintaining long-range structural consistency, particularly within generative AI architectures prone to "ontological drift". This paper introduces WorldPT, a novel framework and dataset that formalizes worldbuilding through Directed Multilayer Attributed Graphs. By implementing a Grounding Directionality Axiom and a hexapartite layering system (Structural, Causal, Temporal, Social, Ontological, and Symbolic), we transition from unstructured text-centric models to machine-verifiable narrative structures. The dataset is uniquely curated in Portuguese, aiming to democratize access to advanced computational narratology resources for the Lusophone community. To evaluate the framework, we applied Social Network Analysis (SNA) metrics to a case study of Tolkien's Middle-Earth universe. Results reveal a "Small-World" topology (average path length of 2.68) and a predominant structural layer (48.7% of connections), quantitatively fingerprinting the setting as a structural-driven worldbuilding. Furthermore, we propose the Cross-Layer Coupling (CLC) metric to identify "lore-shifters" entities whose multidimensionality transcends individual layers. Our findings demonstrate that WorldPT provides a robust foundation for building ontologically stable and interconnected narrative experiences, bridging the gap between graph-based knowledge representation and creative storytelling.

Abstract: This study examines whether the monthly inflation rate of the energy-related consumer price index component in Turkey can be nowcast more accurately with mixed-frequency indicators. An expanding-window pseudo-out-of-sample design is used to compare a seasonal naive benchmark with Elastic Net, XGBoost, and LightGBM. The predictor set combines monthly macroeconomic indicators with features derived from daily Brent oil prices, daily USD/TRY exchange rates, and hourly EPİAŞ day-ahead electricity market data for 2012–2025. Forecast performance is evaluated with root mean squared error, mean absolute error, and symmetric mean absolute percentage error, while core-sample forecast differentials are assessed with the Diebold–Mariano test. All machine-learning models outperform the benchmark, and the lowest forecast errors are obtained from the core XGBoost specification. Explainability results from standardized Elastic Net coefficients and SHAP decompositions show that headline inflation and EPİAŞ variables provide the largest share of predictive content, Brent forms a secondary cost channel, inflation expectations are supportive, and exchange-rate variables do not emerge as an independently dominant block. The results support mixed-frequency machine learning as a useful framework for short-run monitoring of energy-related inflation in Turkey.

Abstract: Large language model (LLM)-driven computer use agents (CUAs) automate graphical user interface (GUI) tasks but often re-solve previously encountered subtasks, increasing token use, latency, and instability. We address this limitation with a directed graph-based persistent memory in which nodes represent observable GUI states and edges encode executable action sequences. We formalize the memory-augmented agent as S=〈A,Σ,G,δ,π,Φ〉, define stability conditions by analogy with functional stability theory, and derive token-cost efficiency bounds. In control-theoretic terms, the Manager–Worker architecture becomes a closed-loop system where memory provides experience-based feedback, and selecting between memory retrieval and fresh LLM planning is treated as adaptive control. Experiments on OSWorld show that the proposed agent cuts both LLM token consumption and execution time by about 50% versus a memoryless baseline while preserving comparable success rates (≈36.9% on 15-step and ≈46.9% on 50-step tasks). Structured graph memory therefore improves robustness under perturbation and supports convergent efficiency gains over time.

Abstract: When a large language model (LLM) answers a question using retrieved documents, a well known technique called retrieval-augmented generation (RAG) is used most of the time, while retrieving more documents improves answer accuracy but increases cost and response time, on the other hand retrieving fewer documents saves resources but may miss critical information. Most existing RAG systems sidestep this dilemma by applying the same retrieval setting to every query, regardless of how simple or complex the question actually is. This wastes budget on easy questions and under-serves hard ones. This paper introduces Cost-Aware RAG (CA-RAG), a routing framework that solves this problem by treating each query individually. For every incoming question, CA-RAG selects the most suitable retrieval strategy from a fixed menu of four options: Starts from no retrieval at all to fetching the top document-k = 10 most relevant documents. The selection is driven by a scoring formula that balances expected answer quality against predicted cost and response time. The weights in this formula act as dials: adjusting them shifts the system toward speed, savings, or quality without any retraining. CA-RAG is built on Facebook AI Similarity Search (FAISS) for document retrieval and the OpenAI chat and embedding application programming interfaces (APIs). We evaluate CA-RAG on a benchmark of 28 queries. The router intelligently assigns different strategies to different queries, resulting in 26% fewer billed tokens compared to always using heavy retrieval, and 34\% lower response time compared to always answering directly without retrieval with excellent answer quality in both cases. Further analysis shows that most savings come from simpler queries, where heavy retrieval was never necessary to begin with. All results are reproducible from logged comma-separated values (CSV) files. CA-RAG demonstrates that a small but well-designed set of retrieval strategies combined with lightweight per-query routing can meaningfully reduce the cost and latency of LLM deployments without compromising the quality of answers.

Abstract: Accurate, high-throughput quantification of rice panicles plays a vital role in advancing precision yield prediction. However, transitioning to real-time, edge-deployable unmanned aerial vehicle phenotyping is often impeded by extreme spatial scale variations from altitude fluctuations and complex unstructured backgrounds. To address this, we constructed a comprehensive composite dataset specifically capturing multi-altitude and varying illumination field conditions. We then propose Panicle-DETR, a highly optimized precision phenotyping framework incorporating a frequency-aware CSP backbone. By projecting visual perception into the frequency domain, the architecture inherently suppresses low-frequency environmental noise and minimizes computational redundancy. Furthermore, a Lossless Feature Encoder prevents the irreversible pixel decimation of micro-targets across varying operational altitudes, while a composite metric loss explicitly disentangles heavily adhered panicle clusters. Evaluated on our composite dataset, Panicle-DETR achieved an outstanding detection Precision of 90.97% alongside robust agronomic counting stability, demonstrated by a Mean Absolute Error of 4.28 and an \( R^2 \) of 0.957. With a compact footprint of only 13.78 M parameters, this framework fundamentally overcomes the computational and spatial limitations of traditional vision models, establishing a highly reliable paradigm for autonomous, onboard agricultural monitoring.

Abstract: CAPTCHA systems remain a widely deployed defense against automated abuse, but advances in machine learning have reduced the effectiveness of traditional challenge-based designs and exposed limitations in proprietary risk-scoring systems. This paper presents an adaptive, reinforcement learning-based CAPTCHA defense framework for high-security web applications. The proposed system formulates bot detection as a partially observable Markov decision process and uses a Proximal Policy Optimization agent with Long Short-Term Memory to analyze streamed behavioral telemetry, including mouse movements, clicks, keystrokes, and scrolling, over sequential interaction windows. Based on accumulated evidence, the agent can continue observing, deploy a honeypot, issue graded CAPTCHA challenges, allow a session, or block it. To complement the sequential agent, the framework also includes an XGBoost classifier that produces a session-level human-likelihood score as a supervised benchmark. Experiments on a simulated ticket-purchasing web application using human-generated sessions and multiple bot tiers, including scripted, replay-based, and LLM-powered agents, show strong preliminary performance. Among the evaluated reinforcement learning variants, Soft PPO achieved the best test performance with two reward structures, with one it reached 98.8% accuracy, 100% precision, and 0.987 F1 score, while with the revised reward structure it reached 96.4% accuracy, 100% precision, and 0.963 F1 score. The XGBoost classifier achieved 99.48% accuracy, 1.000 ROC-AUC, and 0.9919 F1 score. The results indicate that sequential reinforcement learning can support accurate and low-friction bot detection, while the accompanying classifier provides an interpretable and efficient benchmark. Compared with proprietary systems such as Google reCAPTCHA v3, the proposed framework emphasizes transparency, auditability, and explicit sequential decision-making rather than black-box risk scoring. Overall, this work introduces an open and adaptive CAPTCHA-defense framework that offers a promising alternative for studying and deploying behavior-based bot mitigation.

Abstract: Large Language Model (LLM) agents have demonstrated remarkable abilities to interact with external tools, functions, Model Context Protocol (MCP) servers, agents, and to take action on behalf of the user. Due to the fast-paced nature of the industry, existing literature does not accurately represent the current state of tool and agent selection. Furthermore, tool and agent selection in production has nuanced components not covered in experimental research. This work provides the first detailed examination of tool selection from a production perspective, distinguishing between the frontend layer where users interact with agents through buttons, slash commands, or natural language and the backend layer where retrieval, execution, orchestration, context engineering, and memory enable scalable reasoning. The paper contributes a unified taxonomy of modern tool and agent selection approaches spanning manual, UI-driven, retrieval-based, and autonomous methods. The backend covers dynamic tool retrieval, chunking, advanced RAG methods, context engineering, reinforcement learning, tool execution, human-in-the-loop processes, authentication, authorization, multi-turn tool calling, short- and long-term memory for tools, and evaluation. Finally, the paper identifies challenges in production components of both the backend and frontend and outlines promising avenues for research and development.

Abstract: Global Positioning System (GPS) spoofing poses severe threats to navigation safety, necessitating robust detection mechanisms with enhanced interpretability. This study proposes Stack-TabNet, a novel stacked ensemble learning framework integrating XGBoost, Random Forest, and the attentive transformer-based TabNet network. To address model opacity, an interpretable feature attribution mechanism is employed to quantify feature contributions and guide optimization. Experiments are conducted on a complex dataset comprising authentic and spoofed GPS signals across four classes, characterized by high-dimensional signal metrics and severe class imbalance. The initial model utilizing all available features demonstrates robust detection capability. Subsequently, an optimized variant utilizes a subset of top-ranked features identified by the interpretation mechanism, yielding further improved accuracy. Comparative analysis confirms that the proposed framework surpasses all traditional machine learning and deep learning baselines. The analysis identifies Pseudorange and Time of Code Delay as the most discriminative features. These results indicate that combining ensemble learning with interpretable feature selection significantly enhances detection accuracy and training efficiency for GPS anti-spoofing applications.

Abstract: This study addresses the problem of zero-shot generalization (ZSG) in deep reinforcement learning by proposing an MNK game strategy learning method based on a Fully Convolutional Deep Q-Network (FCN-DQN). Research in deep reinforcement learning aims to develop algorithms that can generalize well to unseen environments at deployment time, thereby avoiding overfitting to the training environment. Solving this problem is crucial for real-world applications, where environments are diverse, dynamic, and inherently unpredictable. By constructing a fully convolutional reinforcement learning policy network composed entirely of convolutional layers with padding to preserve feature map dimensions, the proposed model is able to handle input boards of varying spatial sizes. The model effectively learns local pattern-based strategies and approximations of the k-in-a-row evaluation function rather than performing global search. Furthermore, due to parameter sharing, the network has a relatively small number of parameters and is able to share policy representations across different board scales, thereby improving both sample efficiency and inference efficiency. Experimental results demonstrate that, after being trained on a 3×3 board, the proposed model is able to achieve a certain degree of zero-shot generalization performance in larger, unseen board environments.

Abstract: Today, data are no longer confined to numerical values arranged in row-by-column matrices or stored neatly within relational databases. One of the defining characteristics of big data is its high variety, encompassing unstructured and multimodal forms such as text, audio, images, and video. These data types dominate contemporary domains including social media, digital humanities, biomedical research, education, and surveillance systems, yet they remain difficult to manage and analyze using traditional data management architectures. To cope with this shift, modern data management systems must move beyond schema-driven designs and incorporate multimodal artificial intelligence capable of understanding, integrating, and reasoning across heterogeneous data modalities. This article examines how multimodal AI—particularly large multimodal foundation models—can be leveraged to support the ingestion, representation, organization, and analysis of unstructured data. It discusses emerging multimodal data management frameworks, outlines a conceptual pipeline for multimodal data analysis, and highlights key challenges related to scalability, interpretability, and governance. By situating multimodal AI at the core of data management, this work argues that effective data analysis in the era of big data requires systems that treat meaning, context, and cross-modal relationships as first-class computational objects rather than afterthoughts.

Abstract: Citizen-reporting platforms generate high-volume, multilingual streams of service requests, yet operational triage often relies on coarse category labels and manual inspection. This study develops an explainable, calibration-aware analytics pipeline for FixMyStreet Brussels reports, combining text-based urgency modeling, topic discovery, and spatio-temporal hotspot scoring to support municipal decision-making. From 522,132 raw reports, we build an English-normalized text field for modeling, derive resolution-time outcomes from closed cases, and curate a 1,000-item gold standard with an explicit high-urgency class. A TF–IDF logistic regression baseline achieves strong classification performance and, after probability calibration, yields well-behaved confidence estimates suitable for risk-aware prioritization. Topic-level analyses reveal dominant themes related to sidewalks, road damage, and bulky waste, and hotspot scores highlight persistent, high-impact issue clusters. Event detection on aggregated signals did not identify statistically significant shocks during the analysis window, suggesting that the observed dynamics are driven by chronic, recurring problems rather than abrupt anomalies. Explainability audits via SHAP expose linguistically intuitive drivers for urgent cases (e.g., dangerous, risk, accident) and complaint-oriented terms (e.g., abandoned, illegal, dirty), providing transparent hooks for governance review.

Abstract: Decentralized federated learning (DFL) eliminates the single point of failure inherent in server-based architectures, enabling peer-to-peer collaborative model training. However, the absence of a central authority makes DFL particularly vulnerable to Byzantine attacks from malicious participants. Existing Byzantine-robust methods often fail to exploit the network topology structure of DFL. We propose TrustGraph-DFL, a novel defense mechanism that leverages graph-based trust modeling for Byzantine resilience. Our key insight is that consistency between a neighbor's model update direction and a node's local validation gradient can serve as an effective trust indicator. Each node computes consistency scores by comparing received updates against locally computed validation gradients, then maps these scores to dynamic edge weights for robust weighted aggregation. Experiments on CIFAR-10 demonstrate that TrustGraph-DFL achieves 3--5% higher accuracy than existing methods under 30% Byzantine nodes while maintaining a low false positive rate (approximately 9% at 50% Byzantine fraction, compared to 35% for Krum).

Abstract: This study examines the multifaceted application of machine learning and artificial intelligence (AI) in two key, dynamically developing sectors: cryptocurrency market capitalisation forecasting and customer service optimisation. An analysis of the effectiveness of various regression models (Linear, Lasso, and Decision Tree Regression) in predicting the market capitalisation of 3 leading cryptocurrencies shows that a model's success is highly dependent on the specific characteristics of the asset. While linear models achieve exceptional accuracy (R2>0.99) for most major and liquid cryptocurrencies, nonlinear approaches like Decision Tree Regression prove superior for assets with more complex and nonlinear market dynamics, highlighting the need for a flexible approach to model selection. In parallel, the study analyses the implementation of AI in customer service, reviewing chat communication data with the AI assistant "Naomi" (January 26–February 8, 2025). The AI "Naomi" demonstrated high overall effectiveness in chat communication, resolving over 60% of inquiries. However, a significant number of unresolved chats due to customer inactivity or AI limitations indicate areas for further optimisation. In conclusion, the effective application of AI and machine learning requires a strategic approach tailored to the specific field. The key to success lies in careful model selection, prioritising technical reliability, and continuous adaptation and optimisation based on empirical data and a deep understanding of AI's limitations.

Abstract: Early identification of Autism Spectrum Disorder (ASD) traits in infants is crucial for early intervention, which can greatly improve the child’s quality of life. Solutions that use voice analysis offer a promising non-invasive way to detect ASD. However, most current studies depend on extracting specific voice markers from certain datasets and do not include validation across different groups. In this paper, we propose a supervised contrastive learning method for identifying ASD based on infant vocalizations. We extend the Time-Frequency Consistency (TF-C) framework from self-supervised learning to a contrastive approach that uses labels. Our method takes advantage of both time-related and frequency-related data through a dual-branch encoder. It applies supervised contrastive constraints during pre-training to reduce variation within classes while boosting separation between different classes in the embedding space. We pre-train the model using diagnostic labels on a dataset that includes typically developing (TD), Attention-Deficit Hyperactivity Disorder (ADHD), and ASD infants from an open-access dataset, and then fine-tune it with a simple classification head. Evaluation on a cross-cohort group of participants shows the model generalizes well and can distinguish ASD from non-ASD infants, achieving up to 100.00 % accuracy on non-verbal vocalizations.

Abstract: The increasing demand for sustainable and affordable smart-city infrastructure has intensified the need for low-cost, near-real-time water-quality monitoring systems. In this study, we propose Water-QI, a low-cost Internet of Things (IoT)-based environmental monitoring platform that combines budget-friendly sensors with deep learning for Water Quality Index (WQI) assessment and forecasting. The sensing platform measures five key physicochemical parameters, namely temperature, total dissolved solids (TDS), pH, turbidity, and electrical conductivity, enabling continuous multi-parameter monitoring in urban water environments. To model temporal variations in water quality under both cloud-based and edge-oriented deployment scenarios, we evaluate multiple Gated Recurrent Unit (GRU) architectures with different widths and depths. Experiments are conducted at two temporal resolutions, hourly and minute-level, in order to examine the trade-off between predictive accuracy and computational cost. In the hourly scenario, the single-layer GRU with 64 units achieved the best overall balance, reaching a validation RMSE of 0.0281 and a test R² of 0.9820, while deeper stacked GRU models degraded performance substantially. In the minute-resolution scenario, shallow wider GRU models produced the best results, with the single-layer GRU with 512 units attaining the lowest validation RMSE (0.025548) and the 256-unit variant achieving nearly identical accuracy with much lower inference cost. The results show that increasing model width can yield marginal improvements at high temporal granularity, whereas excessive recurrent depth consistently harms convergence and generalization. Overall, the findings indicate that shallow GRU architectures provide the most practical solution for accurate, low-cost, and scalable near-real-time water-quality forecasting. In particular, the 64-unit GRU is the most suitable choice for hourly low-complexity operation, while the 256-unit GRU offers the best speed--accuracy trade-off for minute-level edge inference on resource-constrained devices.

Abstract: This study develops a mobile-supported system that local governments can use in their irregular waste collection services within the scope of smart cities. Irregular waste refers to waste that individuals or organizations produce non-periodically, which arises unexpectedly or in an unusual manner. This waste can accumulate within the city and cause environmental pollution if it is not notified to the municipality or local government for collection. Unlike small-volume household waste collected at routine times, irregular waste is generally large-volume waste such as construction rubble, vegetable oil, mineral oil, and garden waste. Municipalities have different collection vehicles with varying capacities to suit different waste types and quantities. To increase efficiency in the waste collection process, waste locations should be sequenced and vehicles appropriate to the waste type should be allocated in planning. In the irregular waste collection system developed in this study, waste locations are marked on the map applications running on mobile devices, and notifications are sent to the municipality. This provides a faster, more traceable, and effortless service compared to traditional telephone or petition-based notification methods. The Google Maps API was used for processing and visualizing the notification locations on the map. Notification data is recorded in a database by municipality, and daily or 4-hour planning is done using this data. In this study, genetic algorithm and differential evolution algorithm were used for vehicle routing and vehicle type optimization. To compare the efficiency of both methods, 4 different scenarios were designed with different numbers of waste locations and different types and quantities of waste, and the successes of the methods were compared. Route optimization is calculated not statically, however, using real-time traffic data with Google Distance Matrix API integration, generating the shortest and most economical travel route between waste locations. In this way, efficiency is increased for densely populated city centers while providing citizens with an innovative irregular waste collection infrastructure using more up-to-date technologies.

Abstract: Processing massive volumes of historical offline data through complex, heavy-weight service applications presents a significant engineering challenge. Traditional batch processing methods often require refactoring sophisticated production logic, leading to code duplication and maintenance overhead. This paper proposes a methodology for large-scale offline data processing utilizing a Kappa Architecture. By treating offline data warehouses (DW) as streaming sources and wrapping complexservice applications as asynchronous consumers, we enable high- throughput processing without rewriting core application logic. We compare this approach against Spark-native and micro- batch paradigms, and show that it better preserves logical paritywhile maintaining engineering velocity. We also present a multi- region intelligent job dispatcher that improves availability and throughput.