Assignment 2

Baseline Results

Increasing the number of iterations

Looking at the evolution of the F1 score and multi logloss over the iterations, one can see that they did not reach a plateau yet. Hence, increasing the number of iterations would likely increase the performance of the model

As expected, the scores improve
Looking at the plot above, one can see that the F1 score on the validation data reaches a plateau as the number of iterations increases over 250

Tuning the learning rate

The chart below shows that the best learning rate value is the default one, namely 0.1

Tuning the boosting type

Again, the default value ("gbdt") reaches the best results

Hyperparameter sweeping (random)

15 runs, 150 iterations each

Retraining the best model for 300 iterations

Conclusions
Unfortunately, after the sweep, the F1 score barely increased in comparison with the model with default parameters ran for 300 iterations (0.0005). Nevertheless, compared to the baseline model, the F1 score is 6% better
Exploring the visualisation from the last sweep, it seems that increasing the number of leaves and the L2 regularization parameter (reg_lambda) could improve the results
The default parameters of the model, especially the learning_rate and boosting_type, work remarkably well for the given data, as shown by the first two sweeps