Journal of Al-Qadisiyah for Computer Science and Mathematics

The task of image captioning, which involves generating descriptive textual content from visual input, is a pivotal challenge in multimodal learning. This research delves into the advancements in image captioning facilitated by Transformer-based models, comparing their performance, architectures, and innovations across various tasks. Traditional models, such as CNNs paired with RNNs, were initially used to extract visual features and generate corresponding captions. However, the introduction of Transformer architectures has significantly enhanced the performance of image captioning systems, allowing for more coherent, context-aware, and grammatically correct captions. This paper explores the evolution of Transformer-based models, with a particular focus on the Encoder-Decoder, Vision-Language Fusion, and End-to-End Transformers models. By analyzing state-of-the-art architectures such as ViT, GPT, BLIP, and CoCa, the study demonstrates how these models address long-range dependencies, utilize self-attention mechanisms, and seamlessly integrate vision and language for improved caption generation. Furthermore, the paper evaluates the strengths, challenges, and limitations of these approaches, including issues related to computational complexity, dataset biases, and caption diversity. Ultimately, this study presents a comprehensive comparison of these models, offering insights into future research directions in the field of image captioning.

Epileptic seizure prediction facilitates proactive rather than reactive medical treatment, thereby promoting patient safety and improved living conditions. One of the most prominent tools employed for epileptic seizure prediction purposes is electroencephalography, considering its temporal resolution and ability to detect brain activity. This paper offers an organized and tightly focused review on new intelligent learning methods applied for epileptic seizure prediction via EEG. This review discusses publicly and privately accessible datasets, preprocessing and segmentation, feature spaces, machine learning and deep neural networks, evaluation schemes, and postprocessing. It points out major bottlenecks preventing more effective implementation, including an overdependence on a few common datasets, preprocessing conventions, cross-validation on different patients, and poor interpretation and implementation strategies. This paper proposes future developments on more reliable seizure predictors.

Brain stroke represents the main contributor to both disability and mortality in the world. The precise and immediate detection of this disease is crucial to save life of patients and enable effective intervention. A deep learning framework is developed in this research to detect and diagnosis brain stroke into three main categories: normal, ischemic and hemorrhagic. Four pre-trained deep learning models are leveraged in this study including VGG19, CNN, EfficientNetB0 and ResNeXt-50_32x4d. An augmentation and pre-processing techniques are utilized in this framework to reduce class imbalance and unify wide imaging data, hence improving model generalization. Evaluation on the a well curated dataset reveals perfect performance especially precision and recall metrics with scores of (99%) among different stroke types. ResNeXt-50_32x4d model demonstrates the best performance due its strong and robust architecture. Next is the EfficientNetB0 delivers also good performance for its efficient architecture and a smaller number of parameters. Overall, the evaluation analysis results of diagnosing brain stroke types refer to the superiority of the proposed approach in comparison with the baseline deep learning methods. This study also overcomes challenges such as false negatives in the diagnosis of early ischemic cases and false positives arising from anatomical variations, suggesting solutions that integrate several AI models. To sum up, this framework demonstrates encouraging pathway to enhance diagnosis of brain strokes by efficient, automated and interpretable AI tools.

Early and accurate cardiovascular diseases detection is indispensable in specifying efficient treatment and prohibiting life-threatening complications. Traditional detection schemes that depend on manual interpretation of electrocardiograms (ECGs) are generally subject to inter-observer variability and are time-consuming. In this paper, a deep learning model is proposed for classifying cardiovascular diseases relying on two benchmark publicly available datasets of “twelve-lead ECG images”. In order to improve signal diversity, distinct pre-processes and augmentation are adapted on these datasets of cardiac patients. The proposed deep learning model of Convolutional Neural Network (CNN) encompasses distinct blocks of layers (Extensible and detachable convolutions) that are intended to possess morphological patterns and wider spatial dependencies in ECG images. Furthermore, multiple layers of batch normalization and dropouts were employed for stabilizing training and achieving generalization. Experimental results revealed a superior classification accuracy for the proposed system with augmentation utilizing the two datasets, outperforming existing related systems. These results demonstrated the capability of the proposed system to assist cardiologists in early diagnosis and preventive treatment.

Collaborative filtering is very important in e-marketing systems since it enables users to search through the large product lists and become better buyers. This paper will look into the application of machine learning algorithms for instance K-Nearest Neighbors (KNN) and Decision Tree Classifier in formulating a collaborative filtering for supermarket products like, food, dairy, canned goods, juices and detergents. Three data files were employed; a supermarket products file of five Categories each having ten products, a supermarket manager’s offers file of the same Categories with each set with four products, and a file of past user’s purchases. The methodology employed involves using a user product matrix to capture purchasing behavior, using KNN algorithms to predict user preferences using the distance of Euclidean, also, a decision tree to set up the decision-making rules for products recommendations implemented. The number of the previous user’s purchases is shown in a histogram. The paper compares the KNN algorithm with the decision tree algorithm on recommendation accuracy, precision, recall, F1-score and execution time. The KNN algorithm yielded the best accuracy (0.9651), precision (0.8233), recall (0.8430), F1 score (0.8293) and execution time (0.896). The decision tree algorithm yielded lower accuracy (0.8392), precision (0.2094), recall (0.2111), F1 score (0.2072), and execution time (3.686).

Next word prediction, or language modeling, is a crucial task in natural language processing (NLP) that streamlines the typing process by recommending the subsequent word, hence saving time during conversations. The method of next word prediction is intricate, as the machine must anticipate the user's thoughts. A recurrent neural network (RNN), specifically Long Short-Term Memory (LSTM), comprehends previous text and forecasts subsequent words, aiding the user in sentence construction. This study seeks to employ RNN, specifically LSTM, to forecast the subsequent word in one of the most intricate languages with limited resources (Arabic). The results indicated an accuracy of 97.49% and loss of 0.5155%, suggesting that it is sufficiently effective for predicting subsequent Arabic words.

Healthcare systems have recently undergone a significant digital transformation, driven by the rapid growth of the Internet of Medical Things (IoMT) and smart sensing technologies. The sensing technologies generate continuous, high-speed stream of medical information that require real-time analysis and processing. Healthcare data streams are evolving over time. Recently, medical data segmentation and clustering are considered one of the most important techniques used to enhance IoMT reliability, scalability and to support the online medical decisions. Furthermore, these techniques employ bandwidth optimization by reducing the overhead and transmission delay. To date, several surveys have also been proposed in the literature. However, current challenges such as real-time processing and dynamic maintaining of wide variety of medical data streams, which raise the question of developing intelligent and adaptive analytical systems for use in the medical field. Therefore, we conduct a comprehensive survey on the recent advancements in the segmentation and clustering methods for healthcare data streams. This survey examines healthcare data streams, employing clustering and segmentation techniques to improve diagnostic accuracy and enable early disease prediction.

The rapid growth of data stream applications, such as Internet of Things (IoT) systems, smart environments, and real-time analytics, has intensified the need for learning models capable of adapting to continuously evolving data distributions. The traditional techniques of batch learning presuppose fixed data and as such they find it hard to sustain performance when concept drift occurs where the characteristics of data evolve slowly, suddenly or repeatedly with time. Incremental learning has become one of the most important solutions to this issue because it allows models to keep on updating themselves with new incoming data and retain the information previously learned and does not force them to retrain at a high cost. This survey provides an in-depth overview of incremental learning approaches that are used in the streaming data setting with concept drift. We critically review supervised, unsupervised and semi-supervised machine learning methods and deep learning and hybrid methods. The review studies the fundamental adaptation techniques, such as sliding windows, replay, prototype-based learning, methods of detecting drift, expansion of dynamic architecture, and stability-plasticity balancing. To each category we comment on the underlying mechanisms, strength, weakness, computational efficiency and their ability to adapt to various drift conditions. Overall, 38 more recent studies are critically analyzed and compared in the wide range of application areas, data sets, and indicators of assessment. Some of the challenges that are highlighted in this survey include catastrophic forgetting, scalability, interpretability, and resistance to complex and recurring concept drift. Lastly, we recognize the important research gaps and establish future plans on how to develop coherent, scalable and explainable incremental learning models of actual streaming data systems in the world.

Network intrusion detection helps to prevent cyber-attacks on the current networks. Classical signature-based approaches cannot identify new attacks, which drives application of the anomaly-based ones. This paper suggests a composite framework of anomaly detection, which combines autoencoder, latent-space Isolation Forest, and weighted ensemble, which is specifically implemented to the NSL-KDD dataset. The autoencoder is only trained on mainstream traffic to be able to grasp the distribution of benign traffic, and anomaly scoring is used by calculating reconstruction errors. Latent-space representations are additionally analyzed using an Isolation Forest to increase the distinction between aberrant designs. The best weighting program is adjusted according to a validation set and the ultimate threshold is selected based on Youdens J statistic to adjust the false positive and true positive. As the experimental findings demonstrate, the proposed ensemble method provides the accuracy of 95.53 percent, the macro F1-score (0.9551), and the ROC-AUC (0.9867), which is remarkably better in detecting the objects than the solo-model autoencoder methodologies. The paper verifies that deep representation learning coupled with ensemble-based scoring is effective in terms of network intrusion detection.

Internet of Things (IoT)-based solutions, especially automation systems, are essential for controlling networked devices and creating adaptive and intelligent environments. End-user-developed IoT services/apps interact and share simultaneous access to devices depending on their preferences, thereby increasing safety, security, and correctness issues. System failures or reduced performance might result from the traditional reactive approach to dispute resolution, which identifies and fixes problems after they have already happened. To tackle important issues, IoT-based applications require special tools to check and analyse their design, helping to find problems and conflicts in the instructions and interactions of IoT applications. This paper employs the Event-B formal method to introduce a novel approach for detecting and resolution conflicts in IoT systems by using priority. To systematically identify and resolve any conflicts, formal models of the IoT system are developed and refined using abstraction and refinement techniques. The objective of this paper is to promote the use of formal Event-B modelling and verification during the initial stages of the development of an IoT system to expedite the detection and fixing of problems. The proposed model addresses several types of conflicts, such as resource, functional, priority, and policy conflicts. The Rodin platform automates verification, ensuring that all declared invariants and attributes are preserved during system operation. As an example of our method in action, we will take a look at an ECG IoT system. The model accurately detects possible conflicts in these situations. We verify our method using the Rodin model checking tool, proof obligations and the ProB animator.

In this work, a deep system for automatic image description is presented, which aims to produce fluent, meaningful, and structurally coherent sentences for input images. The proposed architecture is based on an encoder-decoder framework, in which high-level image features are first extracted by an Inception-v3 deep convolutional network and then fed as a compressed image representation to an LSTM-based language decoder to produce a word-by-word sentence. On this basic structure, a voting-based ensemble learning framework is designed, in which three deep paths, including a fully connected (FC) network, a Softmax linear model, and a sequence-oriented LSTM decoder, are trained independently, and the word probability vectors at the output level are combined with a maximum voting mechanism. The evaluation is performed on the standard Flickr8k database and using BLEU-1 to BLEU-4, METEOR, and ROUGE-L metrics. The results show that the best single LSTM model achieves values ​​of 0.64, 0.39, 0.23, and 0.16 for BLEU-1 to BLEU-4, and 0.22 and 0.50 for METEOR and ROUGE-L, respectively, while the Ensemble model improves the values ​​to 0.74, 0.50, 0.35, and 0.22 for BLEU-1 to BLEU-4, 0.475 for METEOR, and 0.55 for ROUGE-L; such that the relative improvements in BLEU-3 and BLEU-4 are 54% and 41%, respectively. The paired t-test also shows that the difference in Ensemble performance with single models is significant at the 95% confidence level, and compared to the existing methods on Flickr8k, competitive results are obtained and, in some measures, superior.

This paper attempts to design and compare an equalizer for a 5G 2×2 MIMO-OFDM system employing a data-driven ANN equalizer for realistic wireless channel impairments. The focus here is whether a data-driven equalizer approach can attain better or a comparable equalizer performance to classic linear techniques: Zero-Forcing (ZF), Minimum Mean Square Error (MMSE), MMSE-Successive Interference Cancellation MMSE-SIC, Decision Feedback Equalizer (DFE) but taking into account both accuracy and complexity. Build a 5G MIMO-OFDM simulator in Python for the physical 3GPP TDL-C and TDL-E and do add hurts like Doppler spread and existent inaccuracies like carrier frequency offset CFO phase noise and outdated or imperfect channel state information CSI. Propose a deep fully connected ANN and train using supervised learning on pairs of OFDMs exchanged using Monte Carlo simulating the wireless channel from the previous step. Vast range of SNR/Doppler/CFO/phase noise. Use the ANN as equalizer. For evaluation use BER/EVM/NMSE/and FLOPs and inference time per OFDM frame. Simulation results show that the ANN equalizer is much better than ZF and DFE. And matches / slightly better than MMSE and MMSE-SIC across varying SNR/Doppler/CFO/and phase noise – even in more severe settings with TDL-E with imperfect CSI. Finally, the ANN once trained has better latency for inference and FLOPs/value performance than classic equalizer results. Concluding that ANN based is a good avenue with a 5G MIMO-OFDM reception in mind.

Microsoft Excel is commonly treated as a simple spreadsheet program that is mainly applied to do simple calculations, and Python has become a common solution to complex data analysis based on the programming language. In this work, a technical performance analysis of Microsoft Excel and Python (Pandas) is provided as an attempt to assess their compatibility with basic data analysis functions. The comparison is done which consists of loading datasets, cleaning data, calculating features, aggregating and searching them. The main performance parameter used is execution time and measurements are taken at every processing phase so as to provide a detailed and equitable evaluation. To ensure a robust assessment, the methodology utilized two distinct transactional datasets: a medium-scale over 500000 records and a large-scale dataset over one million records. A key contribution of this work is the implementation of a search loop algorithm as a stress test, where execution time was measured for three specific scenarios: searching for a record at the beginning, at the end, and for a non-existent value. According to the experimental findings, Excel can efficiently handle all of the analyzed tasks and its execution time is as approximate Python as both systems are able to complete the tasks in an approximate time interval. Despite Python having more accuracy in procedures and measurements of time, more automation, and higher reproducibility due to code-based workflows, Excel has competitive analytical capabilities in interactive processing of data. These results underscore the idea that Microsoft Excel can no longer be viewed as an exclusively computational device, and instead of that, it can be viewed as a technically competent data analysis platform capable of performing similarly to Python on large-scaled analytical workloads.

The categorization of software requirements as functional (FR) and non-functional requirements (NFR) is an important problem in software engineering that is widely performed using manual analysis by domain experts. The process is tedious, prone to errors, and lacks consistency from one project to another and from one organization to the next. New approaches are needed to address this challenge, we present such an approach, a new hybrid approach, which combines pre-trained transformer models to automate the requirements classification process, and machine learning techniques. We propose a novel methodology that utilizes BERT (Bidirectional Encoder Representations from Transformers) for contextual feature extraction, supplemented with multi-head attention mechanisms and bidirectional LSTM layers for capturing sequential dependencies in requirements text. We combine this deep learning architecture using ensemble classifiers (Support Vector Machines and Random Forest) with a weighted voting mechanism. We perform an experimental validation using two of the most common datasets (PROMISE with 625 requirements and PURE with 969 requirements), showing that our approach can achieve 94.2% accuracy and 94.1% F1 score for binary classification and an average of 89.7 F1 score across six categories (Security, Performance, Usability, Reliability, Portability, and Maintainability) for the multi-class NFR categorization task. Statistical significance tests further confirm that our hybrid model substantially outperforms each of the state-the-art approaches, with absolute gain ranging from 2.1% to 14.4% respective to different evaluation criteria. This directly addresses some of the main challenges in automated requirements engineering, and hence a unique, pragmatic approach towards large-scale software development processes.

This study submits a comprehensive comparative analysis of three advanced optimization algorithms applied to the classic Traveling Salesman Problem (TSP), a cornerstone of combinatorial optimization. The chosen algorithms are the Ant Colony Optimization, Particle Swarm Optimization, and Gray Wolf Optimization, were evaluated on nine standard cases from the TSPLIB95 library: att48, berlin52, st70, eil76, brg180, pa561, gr666, pr1002, and pr2392, reflecting varying problem sizes and complexities. Results, such as random variance, best path length, relative error, and mean ± standard deviation, were obtained after each algorithm was executed in 30 independent runs. The findings provide empirical insights into the strengths, limitations, and scalability of each algorithm across different problem sizes. It is worth noting that the ACO and PSO algorithms demonstrate a superior balance between solution accuracy and robustness, making them promising candidates for solving large-scale combinatorial problems. They also highlight the importance of statistical validation and analysis of variance in comparative optimization studies, and provide valuable insights into the suitability of algorithms across various TSP metrics.

In communication and transaction systems, searching is a crucial procedure, and there are several techniques available to maximize data retrieval. These include the well-known techniques for effective searching in sorted datasets: Hash Search, Ternary Search, Exponential Search, and Interpolation Search. By putting these four algorithms into practice in Python and evaluating their search effectiveness on artificial datasets made up of sorted arrays of distinct random integers with values ranging from 1 to , where n varies from 1,000 to 1,000,000 items, this study evaluates the performance of these algorithms. calculated average search durations for existing elements across 100 queries using a pseudo-process technique with timed code execution. Because they are deterministic exact-match searches, all algorithms achieve accuracy, precision, and recall, and performance data include average search time (in seconds). According to the results, Hash Search performs best with near-constant times of roughly microseconds, followed by Interpolation Search at 1-2 microseconds, and Ternary and Exponential Searches, which increase with dataset size and vary from 1-4 microseconds. Though newer deep learning techniques, such as learned indexes, may predict data distributions and offer speedups and 10–100x size reductions in specific database contexts, these classic strategies are still effective for broad use.

Social media sites such as Twitter (X) changed into high-velocity observatories of collective mood with millions of short, informal utterances recording public reactions to events, products, policies, and cultural moments in near real time. Mining the stream for actionable insight calls for approaches that respect two notoriously hard to handle properties of social text: (i) ambiguity — in reality, most posts display mixed or low-intensity affect rather than a single discrete label; and (ii) contextual drift — lexical and topical signals co-evolve with communities and time. Many standard sentiment pipelines requiring each message to be labeled as a single discrete class (positive/negative/neutral) fail to provide adequate behavioral insights for crisis monitoring or policy assessment. We close this gap by presenting a Sentiment-Aware Fuzzy Clustering model that models sentiment as a continuous signal and community mood as overlapping regions instead of disjoint boxes. We assign a polarity score to each post and then discretize the space using Fuzzy C-Means (FCM) to assign partial memberships to a number of emotional groups. This uncertainty-aware representation is more reflective of actual online behavior (i.e. a post could be strongly positive but still exhibit features of the neutral discourse) and serves as a principled basis for downstream interpretation in population scale.

One of the controversial topics that has been raised in recent decades. Several approaches involving deep learning and machine learning was implemented to investing in the field of next word prediction in multiple language. In Arabic contexts, this topic is still challenges and in its early stages which need more investigation researches. this field suffers from lakes of robust model that designed specifically for predicted next Arabic word and high-quality dataset that used to develop this domain. This paper proposed a parallel hybrid model (AraBERT- knowledge graph) that augmented the pre-trained AraBERT model with knowledge graph to significantly enhanced next word prediction in Arabic corpus. The constructed of the proposed model involved integrating the AraBERT’s contextual vectors with entities of embedding of knowledge graph (KG). The SANAD dataset (195k articles) was used train this model. The model achieved an accuracy of 90% and an F1-score of 91% which outperformed several of fine adaption baseline models including AraGPT-2 (84.8) and AraBERT (82.6%). The significant improvement in results indicates that the combining of contextual and knowledge based has a promising direction for advancing several of Arabic language application including Arabic text prediction, understanding, and auto-generation, and other NLP application.

To reduce the risk of impression muscle fatigue in the medical field, sports, and rehabilitation of disorders and it is muscles are a critical neuromuscular phenomenon. Electromyography (EMG) is the most important Bio signal used to detect muscle fatigue. Many studies over the past few years have been conducted to address the challenge of muscle fatigue (detection, recognition, and prediction). This study presents a review of various approaches to build models, and evaluation metrics, and applications for each structure begins with exploring artificial intelligence (AI) methods such as machine learning (ML) and deep learning (DL), as well as hybrid model showing the way of data acquisition (sensors types, techniques, preprocessing, models…) specially noise affected of data collection of each type of power spectrum. Furthermore, this review compares fatigue detection, recognition, and prediction approaches, highlighting their performance, strengths, and limitations. Finally, a discussion of various aspects of bio-signal-based muscle fatigue, with specific applications and descriptions, and analyses of the datasets used in muscle fatigue to address the most trending issues and challenges in applying upper limb muscle fatigue. The synthesis presented here aims to guide future developments toward robust, interpretable, and real-time neuromuscular fatigue monitoring systems.

Unmanned Aerial Vehicles (UAVs), or drones are increasingly used for civilian and military purposes. But their misuse raises serious concerns related to privacy, safety and security making them double-edged weapons. Consequently, there is an urgent need for effective UAV detection systems to mitigate threats posed by unauthorized UAV operations over restricted territories. With rapid advances in deep learning and computer vision, vision-based UAV detection systems have achieved notable progress. However, the existing reviews often lack systematic algorithmic analysis and clear summarization of trends and limitations. Therefore, this review aims to consolidate and summarize recent vision-based UAV detection methods using deep learning, focusing on convolutional neural network (CNN)-based models and to provide actionable directions for future research. Firstly, this study presents the evolution of UAV detection, key challenges and the pros and cons of the technologies used. Next it presents a summary of the recent advances in UAV detection methods utilizing one- or two-stage detectors only; the literature shows a strong dominance of YOLO-based architectures due to their favorable accuracy–speed trade-off and suitability for real-time deployment. It further summarizes commonly reported evaluation metrics (e.g., precision, recall and F1-score). Finally, it systematically reviews public UAV datasets and their characteristics highlighting persistent dataset limitations, including limited diversity in altitude, weather, illumination and environment which contributes to a comprehensive understanding of their characteristics and applicability.

Credential compromise is one of the most widespread security threats, allowing adversaries to bypass traditional authentication measures and impersonate legitimate users. Traditional intrusion detection systems are often based on network-level or macro-behavioral indicators, which can be easily spoofed by an attacker, thus compromising the effectiveness of those mechanisms. This study presents an improved adaptive intrusion detection system to authenticate user behavior based on micro-digital behavioral profiling. It involves the use of timing of keystrokes, micro-mouse, navigation in the application, and interaction rhythm signatures. The proposed system uses a hybrid model consisting of Long Short-Term Memory (LSTM) sequence prediction and an Autoencoder reconstruction network to learn both structural and temporal variation of user behavior. Also, an adaptive learning module (implemented by a replay buffer and a drift-detection mechanism based on Kullback-Leibler divergence) to continually recalibrate the model when authentic user behavior varies. Experimental testing on a controlled set of 42 subjects in multiple sessions shows that the proposed model can achieve 94.8 0.91 F1-score and 0.05 false-positive rate, which outperforms the use of individual models; adaptive learning brings this number down by half in the case of drift. The comparison analysis proves the superiority of the proposed system in the areas of anomaly detection, stability, and real-time performance, which demonstrates the viability of micro-behavior analytics as a high-resolution security layer that can be used as a persistent authentication and identity-based threat detector.

Analyzing medical data to identify clinically relevant patient groups remains a complex task, particularly for hepatitis C virus patients, as traditional clustering methods often struggle to capture heterogeneous patient relationships and may rely on specific geometric assumptions or be constrained by the number of groups to be found. The proposed framework for detecting patient communities combines local similarity representation and global structural analysis. This framework builds a network of patient similarity using the K-Nearest Neighbors (KNN) algorithm, followed by the detection of patient communities through the Louvain algorithm. The proposed method is implemented on two real hepatitis C virus (HCV) datasets: the Egyptian HCV dataset and the HCV (UCI) dataset. The results of the experiments show effective detection of coherent and stable patient communities. The highest modularity values reach 0.740 and 0.805 for the Egyptian and UCI datasets respectively, at a neighborhood value of K=3. In addition, in this context, low distortion values indicate strong cohesion within the detected communities. Overall, the results confirm that graph-based patient community discovery provides a powerful alternative to traditional clustering techniques for medical data analysis. The proposed framework enables reliable and stable discovery of clinically relevant and interpretable latent patient subgroups.

Early diagnosis of skin cancer especially melanoma is a significant issue because of the visual similarities of lesion types, intra class variability, and drawbacks of the manual diagnosis. Over the past years, there is a wide range of artificial intelligence (AI) methods suggested, but the results reported were quite different because of the variations in datasets, preprocessing pipelines, model architectures, and evaluation protocols. This is a systematic review article of 29 peer-reviewed studies published 2015–2025 that were identified in major scientific databases and divided into three key categories, namely traditional machine learning (ML), deep learning (DL), and transfer learning (TL) frameworks. Comparatively, the studies are assessed by aspects of characteristics of the dataset, feature representation strategies, model complexity, diagnostic performance and clinical applicability. The discussion has shown that deep and transfer learning models tend to perform better than traditional techniques of ML in terms of classification accuracy, yet various issues on the class imbalance, data heterogeneity, the ability to perform generalization as well as efficiency of computation still exist. Another critical research gap that has been found in the literature is the lack of a standardized analytical framework to compare the ML, DL, and TL models under standardized assessment procedures taking into account the real-life limitations in clinical deployment. This paper is a comparative perspective on the current intelligent skin cancer diagnostic systems in a structured way and it identifies the future research directions of stronger, scalable and clinically sound AI-based solutions.

Distributed Denial of Service (DDoS) attacks pose a serious threat to network infrastructure, necessitating real-time detection mechanisms that can operate within distributed environments without compromising data privacy. This paper presents a novel neuromorphic federated learning system that combines bio-inspired spiking neural networks (SNNs) with a federated learning architecture for self-ambient DDoS detection within distributed networks. Our approach employs stateful Leaky Integrate-and-Fire (LIF) neurons with sufficient membrane potential dynamics, Poisson rate encoding for handling temporal information, and network-centric communication protocols. The architecture was evaluated on the CIC-DDoS2019 dataset on five federated nodes with uniform data distribution. Experimental findings reveal 96.64% accuracy, 99.83% precision, 93.44% recall, and 99.65% ROC-AUC score. The system demonstrates real-time performance with an average latency of 0.797ms, a P95 latency of 0.859ms, and a throughput of 1,254.67 samples/second, meeting the critical 100ms requirement for real-time intrusion detection. The federated architecture accommodates collaborative learning without centralising sensitive network data, with an overall communication overhead of 108.13 MB over 20 rounds of training. Our neuromorphic solution offers a promising solution to energy-efficient, privacy-preserved DDoS detection in modern distributed network environments.

Research addressed high-dimensional data in a 2³ factorial experiment model using a completely randomized design. The columns of matrix (X) represented the effects of factor levels and their interactions, along with the overall arithmetic mean, while the rows represented the number of experimental observations. The high-dimensional data problem has been addressed by using several methods, comparing them according to various criteria. Penalty methods (Bridge, LASSO, ALASSO, SCAD) were employed to select significant factors in the model. Through simulation and application of statistical criteria, the performance of these methods was compared, with results Lasso and Adaptive Lasso show poor performance in most cases, with high MSE and MAD values. CAD generally falls in between, offering better performance compared to Lasso and Adaptive Lasso, but not as good as Bridge.

The large number of independent variables often causes problems in the accuracy of the multiple linear regression model. This has motivated researchers to find or select the best model using methods such as forward selection, backward elimination, and stepwise regression. However, these methods have become time-consuming and ineffective when dealing with high-dimensional data. Therefore, statistical literature has proposed the Lasso method and its variants to address these issues. Nevertheless, these methods are sensitive to outliers, which has led to the emergence of several robust studies—particularly focusing on robustifying penalized methods for high-dimensional data, in order to achieve effective variable selection and robust parameter estimation. Among these methods is the Trimmed Penalized Least Squares (RPLTS), which aims to make the Lasso approach more robust against outliers appearing in the dependent variable or in the residuals. However, this method remains sensitive to the presence of leverage points. Accordingly, this research aims to weight the RPLTS method to select the best subset of variables by reducing the influence of leverage points and improving the model’s accuracy. Simulations and real data were used to evaluate the efficiency of the proposed method and compare it with the previous method. The comparison was based on several statistical criteria such as variable selection accuracy, sensitivity to influential variables, specificity of non-influential variables, and mean squared error (MSE) of the model. The method that achieves the highest accuracy, sensitivity, and specificity rates, along with the lowest MSE, is considered superior. The analysis of both real and simulated data demonstrated that the proposed method outperforms the previous one in terms of robustness and efficiency.