“Feature Importance” isn’t enough. You furthermore may need to take a look at “Error Contribution” if you need to know which features are useful on your model.

The concept of “feature importance” is widely utilized in machine learning as probably the most basic variety of model explainability. For instance, it’s utilized in Recursive Feature Elimination (RFE), to iteratively drop the least necessary feature of the model.

Nonetheless, there may be a misconception about it.

The proven fact that a feature is significant doesn’t imply that it is useful for the model!

Indeed, after we say that a feature is significant, this simply implies that the feature brings a high contribution to the predictions made by the model. But we must always consider that such contribution could also be flawed.

Take an easy example: a knowledge scientist by accident forgets the Customer ID between its model’s features. The model uses Customer ID as a highly predictive feature. As a consequence, this feature can have a high feature importance even when it is definitely worsening the model, since it cannot work well on unseen data.

To make things clearer, we’ll have to make a distinction between two concepts:

• Prediction Contribution: what a part of the predictions is as a consequence of the feature; that is comparable to feature importance.
• Error Contribution: what a part of the prediction errors is as a consequence of the presence of the feature within the model.

In this text, we’ll see how you can calculate these quantities and how you can use them to get beneficial insights a couple of predictive model (and to enhance it).

Suppose we built a model to predict the income of individuals based on their job, age, and nationality. Now we use the model to make predictions on three people.

Thus, we now have the bottom truth, the model prediction, and the resulting error: