Abstract:
In recent years, decision support systems have become more and more perva-
sive in our society, playing an important role in our everyday life. But these
systems, often called black-box models, are extremely complex and it may be
impossible to understand or explain how they work in a human interpretable
way. This lack of explainability is an issue: ethically because we have to be
sure that our system is fair and reasonable; practically because people tend
to trust more what they understand.
However, substituting black-box model with a more interpretable one in the
process of decision making may be impossible: interpretable model may not
work as well as the original one or training data may be no longer available.
In this thesis we focus on forests of decision trees, which are particular cases
of black-box models. If fact, trees are interpretable models, but forests are
composed by thousand of trees that cooperate to take decisions, making the
final model too complex to comprehend its behavior.
In this work we show that Generalized Additive Models (GAMs) can be
used to explain forests of decision trees with a good level of accuracy. In
fact, GAMs are linear combination of single-features or pair-features mod-
els, called shape functions. Since shape functions can be only one- or two-
dimensional functions, they can be easily visualized and interpreted by user.
At the same time, shape functions can be arbitrarily complex, making GAMs
as powerful as other more complex models.
