No abstract

The significant rise in the frequency and sophistication of cyber-attacks and their diversity necessitated various researchers to develop strong and effective approaches to address recurring cyber threat challenges. This study evaluated the performance of three selected meta-learning models for optimal multi-class detection of cyber-attacks using the University of New South Wales 2015 Network benchmark (UNSW-NB15) Intrusion Dataset. The results of this study show and confirm the ability of the three base models; Naive Bayes, C4.5 Decision Tree, and K-Nearest Neighbor for solving multi-class problems. It further affirms the knack of the duo of feature selection techniques and stacked ensemble learning to optimize ML models' performances. The stacking of the predictions of the information gain base models with Model Decision Tree meta-algorithm recorded the most improved and optimal cyber-attacks detection accuracy and Mattew's correlation Coefficient than the stacking with the Multiple Model Trees (MMT) and Multi Response Linear regression (MLR) Meta algorithms.

Lower Respiratory Tract Infections (LRTIs) are the second and third causes of pediatric patients' death in Nigeria and the United States of America. It is observed from several reviewed literature that the LRTIs accounted for more than a million children morbidity and mortality yearly due to lack of prompt diagnosis or no diagnosis due to a shortage of medical experts and medical facilities in our localities. Intense research is ongoing on applying machine learning (ML) to its clinical diagnosis and reducing its spread in pediatric patients. In this research, K-Nearest Neighbor (KNN), C4.5 Decision Tree, and Naive Bayes' ML algorithms were used to develop three base diagnosis models with Correlation, consistency, and information gain selected feature of the LRTI dataset, Multiple Model Trees (MMT) Meta algorithm is used to combine and improve the diagnoses of all the base models using stacked ensemble. The preliminary diagnosis findings using base models have established that the information gained feature extraction method performed much better than the other two. It, therefore, suffix that the results from this should be used for further processing. All the models built with the reduced feature set recorded improved diagnoses accuracy more than the model built with the whole feature set. The MMT stacked ensemble models recorded an improvement on the diagnosis of LRTIs in Peadiatric, it recorded the highest diagnostic accuracies improvement of 12.80%, 13.52%, and 12.37%, and lowest diagnostic accuracies improvement of 6.37%, 5.22%, and 6.09% with the MMT stacked ensemble models of the Consistency, the Correlation, and the information gain reduced selected feature set respectively. These experimental results show the potential for this approach to deliver a reliable and improved diagnosis of LRTIs. It is recommended to be used to diagnose LRTIs in primary health care centers to reduce its mortality rate.

The proliferation of Malware on computer communication systems posed great security challenges to confidential data stored and other valuable substances across the globe. There have been several attempts in curbing the menace using a signature-based approach and in recent times, machine learning techniques have been extensively explored. This paper proposes a framework combining the exploit of both feature selections based on extra tree and random forest and eight ensemble techniques on five base learners- KNN, Naive Bayes, SVM, Decision Trees, and Logistic Regression. K-Nearest Neighbors returns the highest accuracy of 96.48%, 96.40%, and 87.89% on extra-tree, random forest, and without feature selection (WFS) respectively. Random forest ensemble accuracy on both Feature Selections are the highest with 98.50% and 98.16% on random forest and extra-tree respectively. The Extreme Gradient Boosting Classifier is next on random-forest FS with an accuracy of 98.37% while Voting returns the least detection accuracy of 95.80%. On extra-tree FS, Bagging is next with a detection accuracy of 98.09% while Voting returns the least accuracy of 95.54%. Random Forest has the highest all in seven evaluative measures in both extra tree and random forest feature selection techniques. The study results uncover the tree-based ensemble model is proficient and successful for malware classification.