Tangram Tree Benchmarks

Tangram Tree is a pure Rust implementation of Gradient Boosted Decision Trees. It has the smallest memory footprint of the leading GBDT implementations and achieves state of the art speed and accuracy. Check it out on GitHub.

The flights dataset contains 9 columns. There are 10,000,000 rows in the train dataset and 100,000 rows in the test dataset. The target column is dep_delayed_15min and it is a binary classification task.

Training Time (lower is better)

LibraryDurationv. Tangram
catboost329 sec9.235x
lightgbm43.47 sec1.22x
sklearn90.4 sec2.537x
tangram35.63 sec1x
xgboost46.84 sec1.315x
catboost
lightgbm
sklearn
tangram
xgboost

Memory Usage (lower is better)

LibraryMemoryv. Tangram
catboost7.291 GB5.071x
lightgbm2.612 GB1.817x
sklearn2.78 GB1.933x
tangram1.438 GB1x
xgboost2.88 GB2.003x
catboost
lightgbm
sklearn
tangram
xgboost

AUC (higher is better)

LibraryAUCv. Tangram
catboost0.73570.9414x
lightgbm0.78070.999x
sklearn0.7580.9699x
tangram0.78151x
xgboost0.75780.9697x
catboost
lightgbm
sklearn
tangram
xgboost