In an attempt to classify SAT scores into 12 brackets and using the binary splitting technique for hsize and hsrank, a model was trained and evaluated. The results, as illustrated in the plot below, demonstrate a general improvement in accuracy over time. However, concerns arise regarding overfitting as the validation loss steadily increases.
To mitigate the overfitting observed in the previous classification model, the PCA technique was applied in combination with binary splitting. While overfitting was reduced, the accuracy remained relatively constant, hovering around 60%.
Despite utilizing classification techniques and PyTorch's capabilities, it became evident that data generation plays a pivotal role in enhancing model accuracy. The classification approach, in this case, did not yield the desired results.