Identifying Student Behavior for Improving Online Course Performance with Machine Learning

In this study we investigate the correlation between student behavior and performance in online courses. Based on the web logs and syllabus of a course, we extract features that characterize student behavior. Using machine learning algorithms, we build models to predict performance at end of the period. Furthermore, we identify important behavior and behavior combinations in the models. The result of prediction in three tasks reach 87% accurate on average without using any score related features in the first half of the semester. Figure 11 — Accuracy on 10 different periods in BEHP5000 by decision tree (with/without pruning) and random forest (with/without cross validation) ..... 30 Figure 12 — Average time spent on predicting performance on final exam by decision tree (with/without pruning) and random forest (with/without cross validation). Figure 13 — Accuracy on 10 different periods in BEHP5000 by decision tree (with/without pruning) and random forest (with/without cross validation) ..... 33 Figure 14 — Average time spent on predicting success and failure in the first attempt by decision tree (with/without pruning) and random forest Figure 15 — Accuracy on 10 different periods in BEHP5000 by decision tree (with/without pruning) and random forest (with/without cross validation) ..... 36 Figure 16 — Average time spent on predicting performance on final exam by decision tree (with/without pruning) and random forest (with/without cross validation). Table 8 — Vertical actual classes, lateral was predicted classes, the number of samples in that the upper left and lower right is hit the prediction, the number that the upper right and lower left is off the prediction, and give the true and false, respectively, positive, negative represents whether prediction is true and Table 9 — True positive and false positive rates on 10 different periods in BEHP5000 by decision tree (with/without pruning) and random forest Table 10 — True positive and false positive rates on 10 different periods in BEHP5000 by decision tree (with/without pruning) and random forest Table 11 — True positive and false positive rates on 10 different periods in BEHP5000 by decision tree (with/without pruning) and random forest Table 14 — Detected top 3 of the most important behavior combinations on 10 different periods in BEHP5000 by random forest with k-fold cross-validation in prediction of success and failure in the first attempt in the final test .......... 43 Table 15 — Detected top 3 of the most important behavior combinations on 10 different periods …