Features often interact with each other under model predictions, and therefore interaction analysis can provide meaningful insights for this model.
The predictions of a model can be decomposed into the sum of feature effects and the feature interaction effects. Here, Interactions are defined as the effect on the models' predictions that occurs by varying the features values after considering their individual contributions. For example, a house pricing model can use individual features, such as the number of rooms in the house and whether it has a pool. In addition, this model can predict that houses that have a pool and have at least 5 rooms will experience a price increase. This additional cost would only be added when the two features interact at some value intervals. Finding strong interactions between the features according to the model could provide meaningful insights to its predictions.
One of the interaction methods that can be used is called Feature Interaction Performance (FIP). This method assigns a score to each feature interaction which attempts to capture how does the interaction affects the model’s performance. The sign of the score indicates whether the interaction has a positive or negative contribution to the prediction and the value indicates the interaction strength. By calculating FIP on each feature pair, one can rank the interactions and examine further the most meaningful ones.
In the above example, FIP was used to calculate the interaction between all feature pairs using a given binary classification model. The graph clearly shows that the features ‘ap_hi’ and ‘ap_lo’ have a strong interaction relative to all other interactions in the data. In addition, the interaction has a negative effect on the model’s prediction (probability to predict ‘yes’ in binary classification).