Skip to content

Learning Curves (XpViz)#

This interface is divided into three sections:

  1. At the top left, it displays the learning curves, in particular:

    • the main-loss, related to the training of the model;
    • the explain-loss, related to the training of the model’s explanations;
    • and finally the total-loss, which is the sum of the two.

  2. At the bottom left, for each selected epoch, all performance indicators are displayed.

  3. On the right, the structure of the Decision Graph at the currently selected epoch is shown, in particular at the epoch of the best retained model (best model).

Tip

It is recommended to analyze the structure of the explanation graph once the learning curves have stabilized, in order to make adjustments to the explanation hyperparameters.

You can visualize the state of the graph at the epoch preceding each pruning, to analyze all nodes that are candidates for pruning (nodes framed in red). The possible reasons for their pruning are:

  1. The proportion of individuals in the node is lower than the value defined by the parameter population_pruning_threshold. The proportion of individuals within each node is shown as a percentage (blue circle).

  2. The homogeneity of the node is higher than the minimum threshold defined by the parameter target_homogeneity_pruning_threshold. The homogeneity of the node is indicated by the gauge displayed inside the node.

If your graph contains several nodes representing a small number of individuals, or if it converges to only two leaves, gradually increase the parameter balancing_weight and retrain the model.

If your graph excessively prunes several nodes due to reached homogeneity, gradually increase the parameter target_homogeneity_pruning_threshold and retrain the model.

In general, it is recommended to perform pruning halfway through the training process. If the parameter prune_step allows for multiple pruning operations during training, study the structure of the graphs at the epochs before and after pruning to adjust the optimal value of the prune_step parameter.