Investigating the nature of interaction between two features and its effect on a model’s prediction might provide insightful knowledge about both the data and the model.
As explained in Features-Interaction, interactions captured the residual effect on a model's prediction of two feature (or more) feature values, after considering the individual effect.
The interaction strength can be conveniently displayed on a graph where the axis corresponds to the features values. The interaction strength is denoted in the graph as the color of each point. This heatmap representation allows the user to gain insight by visually inspect the interaction of the features and their effects on the model’s prediction.
One method to produce this visualization is called Accumulated Local Effects (ALE for short). This method displays the average effect of the combined features values relative to the mean prediction. In addition, the graph only counts the interaction effect and disregard the individual effects of each feature.
In the above example, ALE was used to calculate the interaction between two feature called ‘ap_hi’ and ‘ap_lo’ using a given binary classification model. The graph clearly shows that there is strong interaction between the two features when ‘ap_hi’ has a low value (lower than 120) and ‘ap_lo’ is approximately at value 90 (which is considered high value). Sample instances that falls in that region will get a “bonus” by the model because of the interaction effect.
#explainableai #interactions
Comentários