Node samples
Help

{{ selectedNode.samples[0] }} {{ 'sample' | pluralize: selectedNode.samples[0] }} in the node ({{ toFixedIfNeeded(selectedNode.samples[1], 2, true) }}% of all the samples).

Defined by
↳ 
{{ rule.left }} 
{{ rule.middle }} 
{{ rule.right }}
Show {{ leftPanel.displayRule ? 'less' : 'more' }}

{{ selectedNode.local_error[1] }} wrong {{ 'prediction' | pluralize: selectedNode.local_error[1] }} ({{ toFixedIfNeeded(selectedNode.local_error[0]*100, 2, true) }}% of the node samples).

They represent {{ toFixedIfNeeded(selectedNode.global_error*100, 2, true) }}% of all the wrong predictions.

Feature distribution
 
#
 {{ feature.name }}
In node samples
In all samples  
#
{{ feature.name }}
No values
Chart data not yet loaded
How does it work ?
Close help

Model Error Analysis provides insights on features critically correlated with a model's failures. These directional insights can be leveraged to improve model design, enhance data collection, or identify key samples for manual investigation.

A decision tree is trained to predict whether your original model ({{ originalModelName }}) would successfully predict an individual sample, or not. A successful prediction by your original model is assessed when the predicted class is the correct one. A successful prediction by your original model is assessed when the predicted value is within close range to the target value (threshold used: {{ epsilon ? toFixedIfNeeded(epsilon, 3, true) : '...' }}).

The tree is trained on the test set of your original model, meaning its nodes represent subpopulations of these test samples, with similar characteristics. This way, Model Error Analysis highlights the subsets where your original model made most of its mistakes.

Select a node to display more information on its samples. Nodes with a high fraction of the total error and a high local error are likely to be of greatest interest for error analysis.

See the Plugin documentation  for more information.