5.4: Conclusion

Interaction effects exists amongst features in a dataset. By accounting these effects, we can develop a model that reflects the true relationship amongst features. Both the methods SHAP, and interaction plot discussed in the chapter are useful in discovering interaction effect amongst features.

One caveat while looking for interaction effects is that we should also avoid looking for interaction between different higher-order features of the same feature. For example, body weight and the square root of body weight are derived from the same feature. It will be useful to search for the interaction effect between these two features.

For all the 4 datasets, we included EDA features, higher order features and interaction effect features. In all the cases, there is a possibility of underfitting. Some features may contain more signal than others, whereas some features could contain more noise than signal. To reduce noise and find a limited set of useful features, we can remove the less useful features. There are many proven methods and techniques to perform feature selection. We will discuss these methods in the next section.