Properly predicting students' academic performance is crucial for elevating educational outcomes in various disciplines. Through precise performance prediction, schools can quickly pinpoint students facing challenges and provide customized educational materials suited to their specific learning needs. The reliance on teachers' experience to predict students' academic performance has proven to be less accurate and efficient than desired. Consequently, the past decade has witnessed a marked surge in employing machine learning and data mining techniques to forecast students' performance. However, the academic community has yet to agree on the most effective algorithm for predicting academic outcomes. Nonetheless, conducting an analysis and comparison of the existing algorithms in this field remains meaningful. Furthermore, recommendations for selecting an appropriate algorithm will be provided to interested researchers and educators based on their specific requirements. This article reviews the state-of-the-art literature on academic performance predictions using machine learning approaches in recent years. It details the variables analyzed, the algorithms implemented, the datasets utilized, and the evaluation metrics applied to assess model efficacy. What makes this work different is that relevant surveys in the past 10 years are also analyzed and compared, highlighting their contributions and review methods. In addition, we compared the accuracy of various machine learning models using popular open-access datasets and determined the best-performing algorithms among them. Our dataset and source codes are released for future algorithm comparisons and evaluations in this community.