Mapping Course Learning Outcomes (CLOs) to Program Learning Outcomes (PLOs) is a key requirement of Outcome-Based Education (OBE). However, manual CLO–PLO mapping is time-consuming, dependent on expert judgement, and may lead to inconsistencies. This study investigates automated CLO–PLO mapping using conventional machine learning, deep learning, and transformer-based classification approaches. The problem is evaluated at two levels: binary classification to determine whether a CLO is mapped to a PLO, and four-class classification to predict mapping strength on a scale of 0–3. While binary classification provides an initial assessment of relationship existence, the four-class task is the primary focus because it captures the degree of CLO–PLO alignment required for practical outcome mapping. Sentence-BERT embeddings are evaluated with five machine learning classifiers, while CNN, LSTM, GRU, and hybrid CNN–LSTM architectures are used for deep learning-based prediction. In addition, a class-weighted DistilBERT model is included as an end-to-end transformer-based baseline to address the observed class imbalance. Models are evaluated through repeated experiments using accuracy, precision, recall, F1-score, class-wise and PLO-wise analysis, and statistical significance testing. The results show that binary CLO–PLO mapping can be predicted more effectively than the more challenging four-class mapping-strength task. Performance in the four-class setting varies across mapping-strength levels and PLOs, with minority and semantically ambiguous classes remaining more difficult to predict. Overall, the study provides a comparative evaluation of conventional machine learning, deep learning, and transformer-based approaches, offering evidence on their relative strengths and limitations for developing more consistent and scalable CLO–PLO mapping systems.
The growing complexity of contemporary network infrastructures has contributed greatly to the size of the attack surface of cyber threats, and as a result, the traditional security mechanisms do not aid in the identification of advanced attacks. Intrusion Detection Systems (IDS) have become common tools that are used to monitor network traffic and to perform the detection of malicious code. The traditional rule-based systems tend to fail to detect an attack that has never been seen before. This paper presents a Deep Learning-Intrusion Detection Framework called Deep Learning-Intrusion Detection Framework (DL-IDF), which is developed to increase the level of cybersecurity in enterprise and cloud-based networks. The suggested architecture is a combination of the Convolutional Neural Networks (CNN) and the Long Short-Term Memory (LSTM) network to study the traffic over the network and identify suspicious actions. CNN yields the network traffic data’s high-level spatial features, whereas LSTM learns temporal correlations and sequential patterns of the changing cyber threats. An additional module featuring feature selection and dimensionality reduction is incorporated to enhance computational efficiency and model performance. To evaluate the performance of the proposed model, benchmark cybersecurity data are used to assess its performance across various attack types. Experimental results demonstrate that the DL-IDF model achieves high detection accuracy and recall while maintaining a low false alarm rate compared with traditional machine learning-based intrusion detection systems. The hybrid deep learning architecture detects threats quicker and monitors network traffic live. The suggested system protects vital digital infrastructure from new cyber threats in dynamic network settings with an effective and intelligent cybersecurity solution. DL-IDF reduces the false positive rate from 0.10 to 0.04, while maintaining a training time of 150–190 s and memory usage of 60–72
The field of machine learning, along with many others, suffers from a sharp increase in data size and the dimensions of its features, which raises processing costs and reduces classification accuracy. To reduce feature dimensionality, feature selection (FS) techniques are essential. These techniques work by eliminating features that might negatively affect the classifier’s performance, such as irrelevant or less valuable information. In this study, we present the ONRBO algorithm, a combination of Opposition-Based Learning (OBL) and the Newton–Raphson-Based Optimizer (NRBO). This algorithm boasts stronger search capabilities than the basic NRBO algorithm. NRBO incorporates a Trap Avoidance Operator (TAO) to avoid entrapment in local optima and the Newton–Raphson Search Rule (NRSR) to speed up convergence. The capacity of NRBO to traverse challenging search areas is significantly enhanced by these processes. This method sometimes shows poor search efficiency and early convergence. The ONRBO algorithm, which employs OBL, was developed to address these problems. Depending on the OBL process, ONRBO is expected to have a better search mechanism and the ability to avoid local optimum. Additionally, it increases the pace of convergence and population diversity. Binary versions of the NRBO and ONRBO techniques, referred to as BNRBO and BONRBO, respectively, were obtained and used along with the k-Nearest Neighbor (k-NN) classifier to choose the best feature subset in addressing FS problems. The performance levels of both BNRBO and BONRBO were rigorously assessed using two additional well-known classifiers, support vector machine and decision tree, within a comparative framework. The capacity of each approach to reduce feature dimensions while maintaining or improving classification accuracy was evaluated. Several studies were conducted using 24 widely available medical datasets, collected from multiple reputable sources, to assess the effectiveness of these methods. In addition, the effectiveness of BONRBO was evaluated using three relevant criteria. Evaluation results show that BONRBO outperformed other k-NN-based FS techniques, such as the basic BNRBO and other binary optimization techniques. Specifically, BONRBO outperformed all other competing binary algorithms in 10 datasets based on classification accuracy, in 11 datasets based on the total number of selected features, and in 7 datasets based on fitness values, out of a total of 24 datasets considered.
This study proposes a water quality classification framework that explains the underlying characteristics influencing the classification performance of machine learning models. The framework prioritizes consistency in the preprocessing pipeline over maximizing predictive accuracy to prevent data leakage during the evaluation of objective performance metrics. By combining standard data preprocessing techniques, the proposed framework improved the overall classification accuracy to 69.82
The rapid advancement of deep learning has enabled the generation of highly realistic deepfake videos while creating major challenges for digital media verification. This study presents a hybrid deepfake detection framework centered on Vision Transformer (ViT) representations, attention-based feature processing, and a lightweight discriminator-inspired classification head. Uniformly sampled facial frames are detected and aligned using MTCNN and then processed by a pretrained ViT backbone. Frame-level fake probabilities are aggregated at the video level, while a Frame Consistency Index (FCI) is retained as an auxiliary measure of temporal variation between consecutive ViT representations. To prevent identity leakage, the primary evaluation is conducted using a subject-disjoint Celeb-DF v2 protocol in which identities in the test partition are excluded from training and validation. On the resulting 220-video test set, the proposed model achieves 87.73
