Skip to content

Analyzing Graph Structure and Learning Curves (XpViz)#

Interface overview#

The interface is organized into four areas:

Top left — Learning curves of the self-explainable model#

This area displays the loss curves and the performance metrics defined by the user, for example:

  • main-loss: loss related to training the deep model (prediction)
  • explain-loss: loss related to training the model explanations
  • total-loss: sum of main-loss and explain-loss

Bottom left — Performance metrics#

For the selected epoch, this section shows all metrics by split (train/val/test, depending on your configuration).

Top right — Decision graph structure (learned or being learned)#

Displays the Decision Graph structure at the selected epoch, in particular the one from the best saved model (Best model).

Groups all hyperparameters that control explanation learning.

Analyzing the graph structure#

Before adjusting explanation-related hyperparameters, it is recommended to:

  • Make sure the learning curves stabilize, and
  • The expected performance is reached.

Once these conditions are met, you can analyze the graph structure using the points below.

Analyzing the Homogeneity Gain of decisions (Nodes)#

  • Situation: some nodes show a very low or even zero homogeneity gain. Action: these nodes are good candidates for pruning, via the interactive Split/Prune interface (feature coming soon).

Analyzing the number of graph leaves#

  • Situation: the number of leaves does not match the expected number of predictive regions (e.g., number of classes in classification, value levels in regression, typical behaviors in forecasting, etc.). Action: adjust the following parameters:

  • Balancing-width: increasing it generally increases the number of leaves

  • Proportion-pruning-threshold: decreasing it generally increases the number of leaves (less pruning)

  • Situation: the graph produces either 2 leaves, or very few leaves, while the rest of the nodes/leaves are sparse. Action: gradually increase Balancing-width until reaching the expected number of leaves.