Figure 4.

3D visualization of grid search result over the investigated space of hyperparameters $n_{est}$, $dept{h}_{max}$ and $n_{leaves}$. Bubble sizes implicate the validation score (the larger, the better); the color indicates the required training time (the less, the better). The optimum configuration of a LightGBM model is at $n_{est}=150$, $dept{h}_{max}=3$, and $n_{leaves}=4$ at a test accuracy of 0.9815 and a training time of $0.624 \mathrm{s}$}.