Predicting students' success is crucial in educational settings to improve academic performance and prevent dropouts. This study aimed to improve student performance prediction by combining advanced machine learning (ML) approaches. Convolutional Neural Networks (CNNs) and attention mechanisms were used for extracting relevant features from student data, such as grades, attendance, and demographic information. Then, these features were fed into different ML models, including Extreme Gradient Boosting (XGB), Random Forest (RF), K-Nearest Neighbors (KNN), Support Vector Machines (SVM), Boosted Decision Tree (BDT), and Naive Bayes (NB). The models were evaluated on multiple datasets, which enhanced the reliability and generalizability of the findings. Experimental results demonstrated that the XGB classifier, integrated with CNN and attention layer, consistently outperformed other models across two different datasets. The CNN + Attention + XGB model attained F1-score, recall, precision, and accuracy values of 1.0 for dataset 1 and 0.97 for dataset 2. This hybrid approach effectively captured complex patterns in student data and made accurate predictions. The study highlighted the potential of hybrid ML techniques for identifying struggling students and enabling targeted interventions for improved academic outcomes.