Iraqi Journal of Data Science

Through the tremendous results achieved by its numerous technologies across a variety of sectors, including scientific, technical, medical, and other fields, artificial intelligence has exceeded expectations and overcome constraints. One area of artificial intelligence called "natural language processing" has demonstrated success in a number of natural language applications, including English, Arabic, German, and other languages. Automatic Natural Language Processing is a technique that uses algorithms to mimic human labor in natural language processing, reducing the time and effort required for an individual to undertake tasks necessary for that processing. The goal of information extraction is to automatically extract a tailored set of data from a massive volume of input text. Many online applications depend substantially on data extraction to function. In this research, we will use Natural Language Processing (NLP) to extract relevant information from an Arabic text and transmit it to the recipient via email. Every system or algorithm for natural language processing is assessed not only for performance, efficacy, and efficiency, but also for the discovery of novel processing techniques applicable to the NLP domain. Three techniques are available for assessing the results: F-measure, precision, and recall. When lexical phrases are used, the information extraction methodology produces very good results; its effectiveness ranged from 86% to 100%, with an average precision of 91.4%, average recall of 90.2%, and average F-measure of 90.7%.

The increasing global prevalence of Type II Diabetes (T2D) demands an advanced predictive model, as conventional machine learning algorithms perform poorly on such complex biomedical datasets. The proposed study aims to develop a trustworthy early prediction model by integrating the Firefly Algorithm (FA) for feature selection with a comparative stacking ensemble technique. The most important results reveal that the FA improved Meta Random Forest classifier achieved the best possible balance between accuracy and efficiency, with 88.68% accuracy, 83.26% sensitivity, and an AUC of 94.37%. It is concluded that the hybrid strategy is an efficient and effective approach to early risk stratification in T2D and addresses an important gap in predictive medicine by combining feature selection optimization and ensemble methods. The hybrid strategy provides a platform for future validation studies across different clinical datasets, enabling proactive, data-driven interventions.

As AI applications expand into new horizons, security issues have become a top priority. Traditional AI systems are not very secure and are vulnerable to attacks and data breaches. The interconnected world we live in requires data to be kept private, protected, and verified, which can be achieved through modern security measures that incorporate cryptographic algorithms. This study evaluates the efficacy of five cryptographic algorithms, ARIA, ZUC, Advanced Encryption Standard (AES), Camellia, and HC-256, regarding key generation duration, encryption and decryption speeds, and memory consumption to identify the optimal solution for practical implementations. Results show HC-256 achieving the fastest encryption and decryption times of 105.6 ms and 75.6 ms, respectively, and the lowest memory usage of 1.28 MB for encryption and 1.38 MB for decryption, proving it is the most efficient in performance-sensitive settings. AES demonstrated balanced performance with moderate encryption and decryption times (116.6 ms and 115.6 ms, respectively) and relatively low memory consumption (6.36 MB), while maintaining strong cryptographic robustness. Camellia demonstrated strong cryptographic robustness according to previous literature; however, it exhibited the lowest computational efficiency in the present experimental evaluation, with the highest encryption and decryption times among the tested algorithms, but was found to yield the lowest efficiency for real-time applications, with the greatest encryption and decryption times of 1886.8 ms and 1937.0 ms, respectively. ARIA and ZUC also exhibit different trade-offs: ZUC shows moderate efficiency, while ARIA requires greater resources.

The problem of hepatitis C is worldwide, with many potential health issues associated with this serious disease (e.g., hepatitis C may lead to liver cirrhosis in some patients). Accurate and early diagnosis of Hepatitis C will improve patient care and possibly lower the death rate from this virus. This paper examined the use of machine learning models to classify patients' hepatitis C stage using laboratory-based clinical data from the hepatitis C dataset on Kaggle, with n=615 subjects. The data was split with 80% for training and 20% for testing. A preprocessing methodology was implemented on the dataset to obtain the best fit for multiple models. For example, missing data were imputed using KNN; categorical variables were encoded as numbers (label encoding); and each feature was scaled (standard scaler). To address class imbalance, we used the Synthetic Minority Over-sampling Technique (SMOTE). The models we implemented were: XGBoost, Random Forest, Decision Tree, Logistic Regression, and Gaussian Naïve Bayes. Our results show that the XGBoost/Synthetic Minority Over-sampling Technique performed best, achieving 95% classification accuracy and a macro F1 score of 0.95, compared to the other models we tested. The results provide additional evidence that combining SMOTE with ensemble learning offers a robust decision-support solution for classifying patients with Hepatitis C in a multiclass setting and an efficient means for early clinical diagnosis.

Unlike previous studies, this document proposes an informed baseline solution to SHE. The proposed initialization significantly reduces the search area and the number of iterations needed to converge, while maintaining a high harmonic elimination precision. In this research, a method for multilayer inverters (MLIs) that is based on an initial solution is presented for SHE. The goal is to resolve the SHE issue with high accuracy while simultaneously reducing the number of iterations by a significant amount. The SHE solution is defined as a set of switching angles; the SHE method is implemented to reduce harmonic distortion. To determine the SHE solution, a comparative analysis was conducted on an eleven-level inverter based on the SHE technique, utilizing the Genetic Algorithm, particle swarm optimization algorithm, and gravity search algorithm. MATLAB/SIMULINK simulations with modulation indices of 0.85 were used to solve the multidimensional SHE problem. An inverter with a single-phase Cascade H-bridge (CHB) configuration and 11 levels is used to implement the algorithm.

Security operations centers (SOCs) today are overwhelmed by a continually growing volume of security alerts from a variety of detection technologies. Analysts typically receive thousands of alerts per shift. Many of these alerts are false positives or low-priority events that cause serious analyst fatigue, delayed incident response, and increased organizational risk. The false-positive rate is typically greater than 50% in business environments. This paper proposes and evaluates ExTriage, a novel Explainable Artificial Intelligence (XAI)-based alert triage framework integrating: (i) a high-performance ensemble classifier combining XGBoost and a multi-layer perceptron; (ii) a SHAP-based global and local explanation engine; (iii) a counterfactual explanation module; and (iv) a RankNet-inspired alert prioritization module. Evaluated on CICIDS-2017, CICIDS-2018, and UNSW-NB15, ExTriage achieves a macro F1-score of 97.89% and AUC-ROC of 0.9963 while reducing false-positive rates by 70.8%. In a controlled study with 24 professional SOC analysts, average triage time was reduced from 47 to 8.2 seconds per alert, with statistically significant improvements in trust calibration, decision confidence, and alert disposition accuracy.