Arabic invoice processing requires not only recognition of generic entities, but also interpretation of their invoice-specific roles and relations. This article extends a YOLOv11–AraBERT active-learning pipeline with a layout-aware structural layer for field classification, key-value linking, and product line-item reconstruction. YOLOv11 localizes text regions in a custom Arabic invoice dataset, OCR converts the detected regions to text, and AraBERT performs sequence labeling with an entity inventory that includes PRODUCT. Layout-AraBERT then combines AraBERT representations, predicted entity types, and normalized two-dimensional bounding-box features to distinguish fields such as invoice date, tax amount, subtotal, total amount, product name, quantity, unit price, and line total. On the authors’ custom dataset, YOLOv11 reaches 93.42% mAP@0.5:0.95 and the final AraBERT active-learning batch reaches 89.36% entity-level F1. PRODUCT extraction reaches 87.77% F1. For the structural stage, Layout-AraBERT reaches 91.2% field accuracy and 88.7% key-value linking F1 under clean OCR conditions; at a 20% synthetic OCR-noise level these values decrease to 81.6% and 78.2%, respectively. Evaluation also includes cross-template generalization, ablation analysis, and annotation-efficiency measurements. The resulting framework provides a reference baseline for layout-aware Arabic invoice understanding on the authors’ custom dataset.
The Internet of Medical Things (IoMT) is transforming healthcare by enabling smart devices to generate vast amounts of sensitive data. Federated learning (FL) allows collaborative model training on distributed IoMT data while preserving privacy, but conventional FL methods face critical challenges in this setting. IoMT devices are often bandwidth-constrained and heterogeneous, making communication efficiency a top priority and the presence of faulty client updates can severely degrade global model performance while coping with highly non-IID data. This work proposes SEFA Sparse & Efficient Federated Aggregation which marries client-side dual compression with server-side two-stage robust aggregation. On the client side, SEFA applies top-k gradient sparsification and an adaptive 8-bit quantization scheme to drastically reduce uplink communication per round while incurring minimal accuracy loss. On the server side, SEFA incorporates a two-stage robust aggregation: first, a Multi-Krum with neighbor selection identifies a subset of updates that are closest to the majority second, a coordinate-wise median is computed over the selected updates to produce a fault-tolerant global model update. This design allows SEFA to achieves comparable accuracy to standard FL in IID settings while maintaining high accuracy under non-IID data with 30% faulty clients, outperforming baseline methods including FedAvg, coordinate-median (FedMedian), single-stage Multi-Krum and a recent IoMT FL framework by a modest margin. SEFA is evaluated on three representative IoMT healthcare datasets Pima Indians Diabetes, Body Performance and Maternal Health Risk using lightweight neural network models demonstrating that SEFA offers a practical and robust solution for deploying federated learning in IoMT environments, with strong fault tolerance, efficiency and minimal performance trade-offs.
Inverse kinematics (IK) is a fundamental problem in robotics that involves computing the joint parameters necessary to achieve the desired position and orientation of a robotic manipulator's end-effector. Traditional methods for solving IK such as Jacobian-based solvers, pseudo-inverse methods, and numerical solvers such as Levenberg-Marquardt and Newton-Raphson are faced with serious challenges ranging from local minimum, singularity, computational cost, and the inability to handle highly redundant systems effectively. To address some of these problems, we proposed Multi-layered Graph Convolutional Neural Network (MGCNN) for 7-DOF (Degree-Of-Freedom) redundant manipulators. The architecture consists of 7 GCN (Graph Convolutional Network) blocks each block consists of a GCN layer, normalization, and activation functions followed by 2 fully connected layers for joint regression. A skip-connection mechanism was used to pool and concatenate outputs from all GCN blocks before being fed into the fully- connected (FC) layers of the network which ensures optimal feature utilization across all layers. The model was trained on mixed data (preplanned trajectories and random poses), where each pose of the robotic arm is represented as a graph with joints and kinematic links representing nodes and edges of the graph respectively. Node features consist of the global end-effector poses and local Denavit-Hartenberg (DH) parameters, enabling the network to learn both global and joint-specific relationships. Experimental results demonstrate strong generalization, a narrow confidence interval relative to Mean Absolute Error (MAE), low error susceptibility, and fast inference. The model achieves an MAE below 0.001 compared to ground truth, proving its robustness and efficiency.
Pedestrian trajectory prediction plays a pivotal role in real-world applications such as autonomous driving, unmanned delivery, and intelligent surveillance. However, existing deep learning approaches still face critical challenges, including mode collapse and the generation of unrealistic trajectories in complex environments. To address these limitations, we propose Phase Fusion Network (PFNet), a novel trajectory prediction framework designed to enhance prediction accuracy in intricate digital media scenarios. PFNet introduces an innovative Graph Encoder (GE) that incorporates a probabilistic modeling strategy to better capture spatial features and pedestrian interactions. To mitigate mode collapse, a common limitation in GAN-based methods, PFNet employs a dual-discriminator mechanism that improves both the realism and diversity of predicted trajectories. Additionally, PFNet adopts a two-phase architecture, where the generation phase strengthens spatial representation and the prediction phase refines temporal consistency. Extensive experiments on standard benchmarks, including ETH, UCY, and the Stanford Drone datasets, demonstrate that PFNet consistently outperforms state-of-the-art methods in terms of both Average Displacement Error (ADE) and Final Displacement Error (FDE).
Retrieval-Augmented Generation (RAG) enhances language model responses by incorporating external knowledge. However, its effectiveness heavily depends on the quality of the retrieved documents. Using a fixed number of retrieved documents K often fails to adapt to varying query complexity, leading either to irrelevant retrievals or to missing crucial evidence. To address this issue, we propose DyRAG, a hybrid retrieval framework that dynamically adjusts K based on query characteristics while maintaining computational efficiency. Our method improves retrieval performance by maximizing relevant information and minimizing noise. We evaluate DyRAG across recommendation, question answering, and fact-checking tasks, where it consistently outperforms fixed- K and hybrid baselines. Compared to traditional approaches, DyRAG demonstrates greater robustness and adaptability across diverse domains.
Background The difficulties with social communication and reserve and repetitive behaviors are the symptoms of neurological disorder, known as autism spectrum disorder (ASD). For prompt intervention and better results, early diagnosis of ASD is essential.Objective The main aim of this research is to diagnose ASD at early stage using a multi modal dataset that include behavioral, social, and eye contact data to forecast the severity of ASD with the help of machine learning algorithm.Method This work investigates the use of particle swarm optimization (PSO) for hyperparameter revising of machine learning models with explainable AI features. One-hot encoding, missing value imputation, and SMOTE were used in the data preprocessing step to solve class imbalance. Hyperparameters were optimized using PSO for Support Vector Machine (SVM). Using stratified k-fold cross-validation, the PSO-optimized models' performance was contrasted with baseline models. The results show that PSO-based hyperparameter tuning greatly improves ASD level prediction recall and accuracy.Result The best result of accuracy 98.60% was obtained by the ensemble model, which combined Random Forest, Gradient Boosting, and PSO-optimized SVM classifiers. This showed how well PSO works to increase the accuracy of ASD diagnoses.Application This study can give to the more reliable and precise ASD diagnostic instruments.
In recent years, neural heuristics leveraging deep reinforcement learning have exhibited considerable promise in addressing multi-objective combinatorial optimization problems (MOCOPs). Nonetheless, challenges persist in attaining both high learning efficiency and optimal solution quality. To address this issue, we propose a novel multi-objective optimization algorithm grounded in information geometry and machine learning principles, which integrates adaptive gradient descent with meta-reinforcement learning techniques to effectively tackle MOCOPs. In this paper, we present a meta-learning framework aimed at enhancing model performance in multi-objective combinatorial optimization through tensor remodeling, preconditioned gradient descent, and entropy regularization strategies. Experimental results demonstrate that the proposed method yields significant performance improvements across several classic multi-objective combinatorial optimization challenges, including the Multi-objective Traveling Salesman Problem (MOTSP), Multi-objective Vehicle Routing Problem (MOCVRP), and Multi-objective Knapsack Problem (MOKP).
Traffic sign detection is a fundamental component of intelligent transportation systems, yet remains challenging due to the small size of signs, visual occlusions, and complex environmental conditions. In this paper, we propose a novel YOLO-based architecture enhanced with multi-scale attention and Transformer modules to address these limitations. Specifically, a Convolutional Block Attention Module (CBAM) is employed to refine spatial and channel-wise features, while a C3 Transformer (C3TR) module introduces multi-head self-attention to capture global contextual information. The proposed enhancements significantly improve the model's ability to detect small and visually degraded traffic signs. Evaluated on the German Traffic Sign Detection Benchmark (GTSDB), our model achieves a mAP@0.5 of 96.75%, mAP@0.5:0.95 of 81.18%, precision of 97.05%, and recall of 95.07%. Compared to YOLOv5 s, this reflects relative gains of +11.2% in mAP@0.5, + 26.6% in mAP@0.5:0.95, + 1.6% in precision, and +20.0% in recall, with a 41.9% reduction in model size. It also outperforms YOLOv8, YOLOv7-tiny, and Faster R-CNN, particularly for degraded signs. For real-time deployment on embedded systems, the model is optimized using NVIDIA TensorRT. This optimization significantly reduces inference latency and computational load while preserving high detection accuracy, making the model well-suited for ADAS and autonomous driving applications.
With the rapid development of AI-driven personalized services, model training increasingly depends on highly sensitive user-side data, such as location, social behaviour, and biometric information. These data not only exhibit pronounced non-independent and identically distributed (non-IID) characteristics but also pose serious privacy risks when processed centrally. Achieving efficient personalized modelling while preserving data locality and privacy has thus become a critical challenge in the evolution of personalized AI. In recent years, personalized federated learning (PFL) has gained significant attention for its strong performance in addressing non-IID data challenges. However, existing approaches often fall short in effectively balancing collaborative efficiency with personalization. To overcome this limitation, we propose FedDAC, a dynamically adaptive, collaboration-enhanced personalized federated learning method. By quantitatively assessing the responsiveness of each parameter to non-IID data, FedDAC dynamically selects collaborative clients, ensuring effective cooperation while retaining personalized feature information. Extensive experiments on four benchmark datasets (EMNIST, CIFAR-10, CIFAR-100, and Tiny ImageNet) under two pathological non-IID settings show that FedDAC consistently outperforms strong baselines. It improves accuracy by 1.5-3.2% on average, reaching 5.9% on highly heterogeneous tasks.
Recently, growing research has focused on analyzing the emotional impact of films, driven by the potential to enhance audience engagement, improve content recommendation systems, and deepen narrative understanding. Advances in artificial intelligence (AI) have opened new avenues for automated emotion recognition across various domains. This study explores the integration of AI, specifically ChatGPT, with human expertise to analyze complex emotional dynamics in cinema. Using a key scene from Sophie's Choice, we investigated ChatGPT's ability to recognize and interpret emotions through a multimodal approach combining visual, auditory, and textual inputs. Two human judges segmented 56 sequences based on characters' facial expressions, and AI tools were employed for emotion analysis of facial recognition, dialogues, audio cues, and cinematographic elements, through an approach that simulated a real-world analytical workflow and the output of several libraries. Results highlight ChatGPT's ability in detecting nuanced emotional cues, such as micro-expressions and contextual elements, while acknowledging challenges in low-visibility conditions. This study demonstrates AI's capability to decode emotions and narrative dynamics and highlights the importance of multimodal frameworks in improving AI-driven emotion recognition. Future research should examine fully autonomous AI analyses and integrate subjective viewer experiences to bridge the gap between AI interpretations and human emotional responses.
A well-structured curriculum is fundamental for providing students with a coherent and meaningful educational journey. However, distributed authorship, informal rules for writing syllabi, and the constant need for updates make curriculum development a highly challenging task. This paper introduces a novel framework that automates the analysis of existing curricula by detecting areas of inconsistency. It utilizes a Large Language Model to extract core concepts and prerequisite relationships directly from unstructured text in course syllabi. To ensure correctness and uniqueness, the extracted entities are linked to Wikidata, a collaborative and general-purpose knowledge graph. Subsequently, a curriculum knowledge graph is constructed based on the relationships between courses and educational concepts, laying the foundation for automated symbolic analysis. We demonstrate the effectiveness of our approach through experiments on the curriculum of the 'Computer Science and Intelligent Systems program offered at AGH University of Krakow. The results are promising, as the tool provides actionable insights on how to improve the curriculum and avoid the most common mistakes.
We present an initial automated test to evaluate LLMs' capacity to perform inductive reasoning tasks. We use the GPT-3.5 and GPT-4 models to create a system which generates Python code as hypotheses for inductive reasoning to transform sequences of the One Dimensional Abstract Reasoning Corpus (1D-ARC) challenge. We experiment with three prompting techniques, namely standard prompting, Chain of Thought (CoT), and direct feedback. We provide results and an analysis of cost-to-success rate and benefit-cost ratio. Our best result is an overall 25% success rate with our CoT prompting on GPT-4, significantly surpassing the standard prompting approach. We assess the programming capabilities of the LLM by analysing the execution rate and errors of the generated code for inductive reasoning. We discuss potential avenues to improve our experiments, testing other strategies, and combining deductive reasoning with LLM-based inductive reasoning.
The improvement of IoT security solutions is nowadays more current and urgent because of the large masses of vulnerabilities, cyber attacks, data theft, and other threats related to the use of IoT devices. Most IoT datasets are imbalanced, as benign traffic dominates while malicious traffic is scarce. Additionally, IoT traffic is sensitive and rarely available for public research. Generating realistic synthetic data is essential for overcoming these limitations. Our study focuses on Deep Learning models for network intrusion detection by implementing Tabular Generative Adversarial Networks (TGAN) to address class imbalance. GANs help by increasing the proportion of rare malware samples, improving model training and detection accuracy. In this paper, we rely on the UNSW-NB15 and NSL-KDD datasets to address the issue of imbalanced classes. We propose a new approach that we called FS-TGAN, which is based on feature selection methods and TGAN model for samples generating. The time is a crucial parameter for security tools such as IDS and antivirus, where a lot of data must be analyzed to look for malware, anomalies, or anything suspicious that might be trying to penetrate the system. For this purpose, we reduced the number of features to eliminate redundant features or those that are highly correlated. The results show that TGAN performs well. We achieved 99.03% of accuracy with the UNSW-NB15 dataset, demonstrating a significantly reduced error rate by learning to provide new unseen data that share the training set's statistics, with a detection time of 0.230 ms per traffic set.
The Self-organizing Map (SOM) is a specialized artificial neural network (ANN) that facilitates data visualization, aiding in the understanding of high-dimensional data and representing clustering mechanisms by grouping similar data together. However, the conventional SOM demonstrates limitations when handling non-Gaussian data distributions, often resulting in suboptimal initial placements, slower convergence, and less effective clustering. To address these challenges, we propose a novel approach to enhance SOM performance by integrating a learning-by-epoch strategy. This strategy introduces a step in the SOM algorithm that checks the kurtosis and skewness of input vectors. If significant deviations are detected, normalization is applied to ensure that the input data falls within an appropriate range. In our experiments using the Iris dataset, the conventional SOM achieved an error rate of 0.0889 after converging in 2100 iterations. The modified SOM reported in previous studies and k-means algorithm yielded error rates of 0.0444 and 0.0476, respectively, with convergence in 1950 iterations. Notably, our proposed SOM outperformed both, achieving an excellent error rate of 0.022 and converging in just 1630 iterations. Additionally, when applied to satellite images, the basic SOM exhibited under-segmentation issues, failing to accurately delineate distinct land cover regions. In contrast, the segmentation results from our modified SOM demonstrated superior performance, yielding a more accurate and finely segmented image. The proposed method achieved the highest Normalized Mutual Information (NMI) across all tested satellite images. Table 1 summarizes these comparative results among different SOM algorithms, highlighting that our proposed modified SOM consistently outperformed other methods in terms of cluster error rate and convergence iterations.
Aspect-based sentiment analysis (ABSA) is a fine-grained analysis of sentiments and opinions expressed in text document which set aside the na & iuml;ve assumption that a document expresses opinion about a single topic or aspect. The main objective of ABSA is to spot, extract and identify the polarity of different entities and aspects in an opinionated document. Based on the previous works, ABSA can be categorized into three subtasks: Aspect-category sentiment analysis (ACSA), Opinion Target Expression Sentiment Analysis (OTESA) and Aspect-term sentiment analysis (ATSA). This research presents an end-to-end multi-task approach to performing the three categories of ABSA on a single pipeline. A ternary multitask learning objectives classifiers were built on top of the baseline spanBERT language model which was originally pretrained for span extraction. The input to the model consists of two merged segments of entity premises and context data hypothesis in a similar passion to reading comprehension downstream task in natural language processing. The ternary downstream tasks were built on the contextualized output embeddings of pretrained spanBERT entangled with cross-layer attention mechanism to associate context with the aspect-term span extraction, aspect sentiment polarity detection and entity-aspect entailment. A span masking approach was also proposed to address multiple-aspects text using an iterative outputs-inputs loopback. The span masking process replaces each word in a previously detected span of text with a special [MASK] character and then feeds back the entire sentence into the encoder input of the model for next run. The technique forces the encoder to look elsewhere for the next span prediction. The loopback span masking terminates when the span classifiers predict a special token [CLS] as the beginning and end of the span signaling the absent of relatable span to be extracted. Experimental results validate the approach as impressive results were obtained outperforming most of the compared research with benchmark ABSA datasets.
Using cross-validation to predict the accuracy of a classifier on unseen data can be done reliably only in the absence of dataset shift, i.e., when the training data and the unseen data are IID. In this work we deal instead with the problem of predicting classifier accuracy on unseen data affected by prior probability shift (PPS), an important type of dataset shift. We propose QuAcc, a method built on top of “quantification” algorithms robust to PPS, i.e., algorithms devised for estimating the prevalence values of the classes in unseen data affected by PPS. QuAcc is based on the idea of viewing the cells of the contingency table (on which classifier accuracy is computed) as classes, and of estimating, via a quantification algorithm, their prevalence values on the unseen data labelled by the classifier. We perform systematic experiments in which we compare the prediction error incurred by QuAcc with that of state-of-the-art classifier accuracy prediction (CAP) methods.
This study presents the development and fine-tuning of an Artificial Intelligence (AI) model based on the BERT model to automate the manual case assignment process at the IT service desk. The manual case assignment process in the university has proven inefficient and error-prone, negatively impacting service quality and user satisfaction. Integrating BERT, a specialized transformer for Natural Language Processing (NLP) and Machine Learning (ML) tasks, has enabled high-precision and efficient automatic case classification. This study used a dataset composed of 33,000 original records and 77,000 artificially generated records. The results demonstrated a marked improvement in the speed and precision of case assignment, allowing human resources to concentrate on other specific tasks. This article provides practical guidance for developing and tuning AI models in IT service management, highlighting the operational benefits and implications of digital transformation in academic institutions.
This paper introduces CLAM, a hybrid deep learning framework that integrates CNNs, LSTMs, and Attention Mechanism (AM) for straightforward multi-step stock trend forecasting. By leveraging CNNs for spatial feature extraction, LSTMs for capturing temporal dependencies, and AM for dynamically focusing on relevant data, CLAM significantly outperforms traditional models in predictive accuracy. Evaluated on diverse stock datasets from different industries, CLAM demonstrates an average reduction of over 80% in MAE and RMSE compared to standalone CNN, LSTM, and fused CNN-LSTM. The model’s ability to capture both short-term and long-term trends is particularly advantageous for real-time financial trading, resulting in 75% trend prediction accuracy, with most cases witnessing consecutive accurate forecasts of flash crashes or uptrends, which aids in strategic investment decisions and risk management. Code and data are available at: https://anonymous.4open.science/r/CNN-LSTM-AM-AB13/src/CLAM.ipynb .
We present a study about the prediction of long-COVID sequelae through multi-label classification (MLC). Data on more than [Formula: see text] patients have been collected during a long-COVID study at Ospedale Maggiore of Novara (Italy), considering their baseline situation, as well as their condition on acute COVID-19 onset. The goal is to predict the presence of specific long-COVID sequelae after a one-year follow-up. To amplify the representativeness of the analysis, we carefully investigated the possibility of both augmenting the dataset by considering situations where different levels in the number of complications could arise, and reducing the number of features to be considered for prediction. In the first case, MLSmote under six different policies of data augmentation has been considered, while in case of feature reduction we have generated new datasets via both a supervised and an unsupervised dimension reduction approach (Relief and PCA respectively). A representative set of MLC approaches has been tested on all the available datasets. Results have been evaluated in terms of Accuracy, Exact match, Hamming score and macro-averaged AUC; they show that MLC methods can actually be useful for the prediction of specific long-COVID sequelae, under the different conditions represented by the different considered datasets. In addition, interpretability of the results has been addressed through an approach based on the SHAP method, showing that clinical interpretations of specific predictions can be actually captured by the method, together with the observation that data augmentation techniques do not harm such a kind of explanations.
