Skip to main content

Table 6 Performance comparison of all models on the training and testing set. The \(R^{2}\) score and root mean squared error (RMSE) indicate how well we can fit the true reading progress. Accuracy measures if the direction of progress is correctly predicted and the f1 score is a combination of precision and recall. Since both, accuracy and f1 score are the same, we know that precision (positive predictive value) and recall (sensitivity) have the same value. For all, except RSME, holds that higher values indicate a better fit. Test and train values are close to one another, which suggests that no strong over-fitting occurred. We finally chose linear regression and ridge regression as final models

From: Both sides of the story: comparing student-level data on reading performance from administrative registers to application generated data from a reading app

Split Metric name Accuracy \(R^{2}\) RMSE f1
Test linear regression 0.621 0.159 16.400 0.621
ridge regression 0.621 0.180 16.197 0.621
extremely randomized trees 0.568 0.114 16.743 0.568
LightGBM 0.636 0.141 16.492 0.636
random forest 0.575 0.086 17.005 0.575
SVR (RBF kernel) 0.594 0.136 16.534 0.594
SVR (linear kernel) 0.620 0.157 16.338 0.620
Train linear regression 0.648 0.158 16.568 0.648
ridge regression 0.642 0.167 16.481 0.642
extremely randomized trees 0.680 0.293 14.968 0.680
LightGBM 0.660 0.197 15.951 0.660
random forest 0.698 0.403 13.751 0.698
SVR (Gaussian kernel) 0.657 0.193 15.990 0.657
SVR (linear kernel) 0.663 0.172 16.198 0.663