How to understand a model
To better interpret machine learning (ML) models, we can use Shapley values.
The concept behind a Shapley value was first developed in the context of Game Theory for games where multiple players strive to maximize a reward or score.
Consider a cooperative game such as a trivia quiz where players form teams and try to answer questions. To identify the strongest players, we must evaluate them in the context of their teams. For example, perhaps Player A makes many useful contributions when on a team with Players B and C, but makes fewer contributions when Player D joins the team, possibly because they have overlapping knowledge. To accurately evaluate Player A’s performance, we look at all possible combinations of teams and so determine how much Player A contributes overall. This is exactly the idea behind Shapley values.
Analogously, we can determine Feature A’s contribution to a ML model from its Shapley value. Here, we look at the change in prediction based on different Feature A values, evaluated for all possible combinations and values of each feature. Having found each feature’s contribution, we can better understand, interpret, and inform on our ML model.