Hi there! I would appreciate any input on my most recent experiment, in which I utilize an XGBoost model to pinpoint the critical characteristics that, for each position, predict whether an NFL player will be drafted. In order to identify the top 5 athletic attributes by position, this project entails thorough data cleansing, exploratory data analysis (EDA), the development of relative performance measurements for talents, and the application of the model.
1 Like
Xgboost isn’t very good for feature importance. I like doing a wide & shallow random forest. Shapley plots from a logistic regression would be good as well.
In comparison to only the logistic regression coefficients, what benefits might Shapley charts offer?
1 Like
I will take a stab at this post. Feature significance does not imply “causality.”
That’s one thing if you’re searching for “the most predictive” characteristics. However, this does not mean that associative measures of predictive utility are the most significant.
1 Like
When I think of how these things are used, the rhetorical differences between “Important”, “predictive”, and “associative” are not very useful for me. “Causal” isn’t ever allowed to enter the discussion.
I mention this here because many people, even practitioners in the space don’t understand the differences sadly