The following partial dependence has been auto-computed after training on the full dataset, without the sample weights. You can recompute it with the sample weights and get the feature distribution. The following partial dependence has been auto-computed after training on the train set, without the sample weights. You can recompute it on the test set with the sample weights, and get the feature distribution.
Partial dependence

Partial dependence plots show the dependence of a model to one of its features,
averaging the other features

Partial dependence not computed yet

Compute now to discover dependence of this model to {{ uiState.selectedFeature }}

{{ nbPoints[uiState.selectedFeature] }} bins for {{ nbPoints[uiState.selectedFeature] }} most frequent modalities of {{ uiState.selectedFeature }}, computed on {{ computedOnStr() }}

Reading tips

A partial dependence plot shows the dependence of the predicted response on a single feature. The x axis displays the value of the selected feature, while the y axis displays the partial dependence.

The value of the partial dependence is by how much the prediction is higher or lower than average.

The value of the partial dependence is by how much the log-odds are higher or lower than those of the average probability.

The value of the partial dependence for a class is by how much the log-odds for this class are higher or lower than those of the average class probability.

Special modalities:
  • {{ labelsRemaping["__DKU_OTHERS__"] }}: This is the aggregation of the modalities that are not displayed (weighted by their distribution)
  • {{ labelsRemaping["__DKU_UNREPRESENTED__"] }}: This emulates a modality that the model doesn't know, it has been clipped by the feature handling or this modality is not present in the train set
  • modality*: This modality is not represented in the model (see above)
  • modality**: This modality is dropped by the preprocessing, no partial dependence can be computed

Note : the log-odds for a probability p are defined as log(p / (1 - p)). They are strictly increasing, ie. higher log odds mean higher probability.

Note: because you are using k-fold cross-testing, the partial dependence is computed on the full dataset.