The XGBoost Vector-Leaf Model
SMRTR summary
XGBoost's new vector-leaf model offers a smarter way to handle multiclass predictions by building one shared tree whose leaves store scores for every class, instead of building separate trees for each class. Tested on ten real-world datasets, vector leaves achieved lower prediction error on nine of ten while producing much smaller models, sometimes just one-ninth the size, making them faster and more resistant to irrelevant features.
SMRTR provides this summary for quick context. The original article belongs to Hacker News.
Read the original article