Khalid Alalawi

Machine learning intrusion detectors are accurate but opaque, so explainable AI justifies their alerts and bias mitigation evens out detection across imbalanced categories. These are usually studied independently, and whether making a detector fairer changes how faithfully it can be explained is unknown. In an interventional design, we train an unmitigated baseline and three imbalance mitigations on a one-dimensional CNN, an FT-Transformer, and XGBoost, applying class weighting and oversampling to all three and class-weighted focal loss to the neural models. Faithfulness is measured with deletion-based comprehensiveness and sufficiency, before and after each mitigation, per category, on CICIoT2023 and CICIDS2017. Bias mitigation affects faithfulness unevenly. On the primary dataset, the rare categories most helped gain recall and more faithful explanations, while the largest penalty fell on a category whose detection scarcely changed. XGBoost, explained by exact TreeSHAP, showed no comprehensiveness cost of its own. Among the neural mitigations, focal loss was the most costly, lowering comprehensiveness by up to 0.13 while reducing the recall parity gap by 0.42–0.45, and oversampling was nearly free; this pattern held on both datasets, more weakly on the second, where class weighting and oversampling were not reliably separated. Fairness and interpretability need not be in tension when the mitigation is chosen well, and faithfulness should be evaluated per category whenever a fairness intervention is applied.

Background: Hepatitis C virus (HCV) affects approximately 50 million people worldwide and progresses silently through distinct hepatic stages, yet most machine learning staging approaches offer no clinician-facing explanation and rarely evaluate whether explanations hold across architectures. Methods: We trained five models—Logistic Regression, Random Forest, XGBoost, LightGBM, and SVM—on the UCI HCV dataset (615 patients, four classes after merging the seven-instance suspect blood-donor group into a single Donor/Control class) and applied SHAP, LIME, and DiCE counterfactuals. A cross-model Spearman agreement analysis quantified feature ranking consistency, with a leakage-controlled pipeline and five-fold stratified cross-validation applied. Results: In the main 5-fold cross-validation, LightGBM achieved the highest macro-F1 (0.684 ± 0.031). Under repeated-seed cross-validation, XGBoost and LightGBM gave closely matched values (0.648 ± 0.023 and 0.645 ± 0.026), indicating comparable robustness among the boosted-tree models, both marginally ahead of the remaining three, with Random Forest close behind. Logistic Regression reached the highest macro-F1 on the single held-out split (0.790), but its cross-validated score was markedly lower (0.598 ± 0.063), underlining how unstable single-split estimates are in a small, imbalanced cohort. SHAP consistently identified AST, GGT, CHE, and ALP across the tree-based models, with CHE emerging as the leading Cirrhosis-stage marker in the boosted models. Cross-model Spearman correlations reached 0.923–0.958 among tree-based models; SHAP-LIME overlap ranged from 1/5 to 4/5. Conclusions: The framework identifies stage-specific biochemical importance patterns consistent with known HCV disease progression. The convergent finding on CHE for Cirrhosis, supported by SHAP in the tree-based models and appearing among the top LIME features in three of the five model-specific Cirrhosis explanations, supports CHE as a candidate marker worth evaluating in advanced disease.

Breast cancer remains one of the most frequently diagnosed cancers worldwide, and improving the accuracy and transparency of automated diagnostic tools is an ongoing clinical priority. This study examines whether three established machine-learning classifiers achieve comparable performance on a standard breast cancer benchmark and whether their SHAP-based feature explanations converge on the same predictive signals across model architectures. Logistic Regression (LR), Support Vector Machines (SVM), and XGBoost were trained and tested on the Wisconsin Diagnostic Breast Cancer (WDBC) dataset using a shared preprocessing pipeline to ensure fair comparison. Hyperparameters were selected through grid search with 5-fold stratified cross-validation, and model performance was estimated using 10-fold stratified cross-validation. Paired t-tests and Wilcoxon signed-rank tests were used to determine whether performance differences between models were statistically meaningful. Shapley Additive explanations (SHAP) values were computed separately for XGBoost using TreeExplainer and for Logistic Regression using LinearExplainer, and Spearman’s rank correlation was used to quantify the agreement between the two models’ feature importance rankings. All three classifiers achieved Receiver Operating Characteristic–Area Under the Curve (ROC-AUCs) above 0.994, with SVM achieving the highest accuracy (0.9737) and F1-score (0.9630). No statistically significant difference was found between any model pair (p > 0.05). The cross-model SHAP analysis yielded a Spearman correlation of r = 0.578 (p = 0.0008), with seven of the ten most important features ranked consistently across both architectures. The agreement between two structurally different models on which features matter most provides evidence that these features carry a consistent predictive signal that goes beyond what any single model’s architecture alone would produce. Cross-model explainability analysis of this kind offers a stronger basis for feature interpretation than the output of any single model.

Learning analytics intervention (LAI) studies aim to identify at-risk students early during an academic term using predictive models and facilitate educators to provide effective interventions to improve educational outcomes. A major impediment to the uptake of LAI is the lack of access to LAI infrastructure by educators to pilot LAI, which typically requires substantial institution-wide efforts and investment to collect related data sets and develop accurate predictive models that identify at-risk students and also provide tools to facilitate interventions. This paper presents a novel LAI framework, termed Student Performance Prediction and Action (SPPA), that facilitates educators to seamlessly provide LAIs in their courses avoiding the need for large-scale institution-wide efforts and investments. Educators develop course-specific predictive models using historical course assessment data. In learning analytics, providing effective interventions is a challenge. SPPA utilises pedagogy principles in course design and interventions to facilitate effective interventions by providing insights into students’ risk levels, gaps in students’ knowledge, and personalised study/revision plans addressing knowledge gaps. SPPA was evaluated in a large undergraduate course on its ability to predict at-risk students and facilitate effective interventions as well as its ease of use by academics. The results are encouraging with high performance of predictive models, facilitating effective interventions leading to significant improved educational outcomes with positive feedback and uptake by academics. With its advantages, SPPA has the potential to catalyse and influence wide-scale adoption in LAIs.

Today, educational institutions produce large amounts of data with the deployment of learning management systems. These large datasets provide an untapped potential to support and enhance decision‐making and operations. In recent times, machine learning (ML) has been applied to develop models utilizing this “big” data to assist in decision‐making. This study presents a systematic literature review into the application of ML to predict student performance. A total of 162 research articles from January 2010 to October 2022 were critically reviewed and analyzed by applying Kitchenham's systematic literature review approach. Our analysis categorized the literature predicting students' academic performance into two categories: (i) predicting student performance in assessments, courses or programs, and identifying students at‐risk of failing their course/program (129 studies); and (ii) predicting student dropout or retention in a course or program (33 studies). Classification is the most commonly used approach for predicting student performance (138 studies), followed by regression (25 studies) and clustering (9 studies). Supervised learning methods are used more often than semi‐supervised learning. Five most popular ML algorithms include the Decision Tree, Random Forest, Naïve Bayes, Artificial Neural Network, and Support Vector Machine. Historical records of students' grades and class performance, academic related data from learning management systems, and students' demographics are the most common features used for predicting students' performance. The most common methods used for feature selection are Information Gain‐based selection algorithms, Correlation‐based feature selection, and Gain Ratio. The general platforms/tools/libraries used in the studies include WEKA, Python, R, Rapid Miner, and MATLAB. We also investigated possible actions considered in the literature to help at‐risk students. We only found very few studies that deployed remedial actions and evaluated their impact on students' performance. In conclusion, ML has shown great potential in the prediction of student performance, but also has many areas of further research.

Predicting student performance and identifying under-performing students early is the first step towards helping students who might have difficulties in meeting learning outcomes of a course resulting in a failing grade. Early detection in this context allows educators to provide appropriate interventions sooner for students facing challenges, which could lead to a higher possibility of success. Machine learning (ML) algorithms can be utilized to create an early warning system that detects students who need assistance and informs both educators and learners about their performance. In this paper, we explore the performance of different ML algorithms for identifying under-performing students in the early stages of an academic term/semester for a selected undergraduate course. First, we attempted to identify students who might fail their course, as a binary classification problem (pass or fail), with several experiments at different times during the semester. Next, we introduced an additional group of students who are at the borderline of failing, resulting in a multiclass classification problem. We were able to identify under-performing students early in the semester using only the first assessment in the course with an accuracy of 95%, and borderline students with an accuracy of 84%. In addition, we introduce a student performance prediction system that allows academics to create ML models and identify under-performing students early on during the academic term.

This paper presents a novel framework aimed at improving educational outcomes in tertiary-level courses. The framework integrates concepts from educational data mining, learning analytics and education research domains. The framework considers the entire life cycle of courses and includes processes and supporting technology artefacts. Well-established pedagogy principles such as Constructive Alignment (CA) and effective feedback principles are incorporated to the framework. Mapping of learning outcomes, assessment tasks and teaching/learning activities using CA enables generating revision/study plans and determining the progress and achievement of students, in addition to assisting with course evaluation. Student performance prediction models are used to identify students at risk of failure early on for interventions. Tools are provided for academics to select student groups for intervention and provide personalised feedback. Feedback reports are generated based on effective feedback principles. Learning analytics dashboards provide information on students' progress and course evaluation. An evaluation of the framework based on a case study and quasi-experimental design on real-world courses is outlined. This research and the framework have the potential to significantly contribute to this important field of study.