Late submission of assignments is one of the major issues in higher education, which leads to poor performance of students and efficiency in teaching. The paper examines the most critical behavioral, academic, and environmental variables related to late submissions amongst the university students in Bangladesh using the Association Rule Mining (ARM) and machine learning algorithms. An analyzed dataset containing 20 attributes and based on a survey comprising 359 valid responses was obtained (out of 361 responses, the rest were duplicates removed). Apriori algorithm found 2,858 association rules that were important, which indicated that irregular attendance during classes and unstable study conditions are the strongest predictors of high probability of late submission (maximum confidence: 90.9%). The 5-fold stratified cross-validation was applied to 6 supervised classifiers SVM, XGBoost, LightGBM, Random Forest, Extra Trees, and Neural Network to ensure severe class imbalance (9.56:1). Extra Trees had the highest cross-validation accuracy (95.96%±2.31%), and XGBoost and Extra Trees had the highest test accuracy (93.06%). Chi-square analysis proved the two variables, attendance at classes and procrastination, to be the most important predictors. On the whole, behavioral variables proved to have a more influential impact than demographic variables. The findings have practical and evidence-based information that can be used in educational institutions to develop specific interventions that can decrease late submissions and enhance academic performance.