MLBenchmarks.jl

This repo provides Julia based benchmarks for ML algo on tabular data. It was developed to support both NeuroTabModels.jl and EvoTrees.jl projects.

Methodology

For each dataset and algo, the following methodology is followed:

• Data is split in three parts: train, eval and test
• A random grid of 16 hyper-parameters is generated
• For each parameter configuration, a model is trained on train data until the evaluation metric tracked against the eval stops improving (early stopping)
• The trained model is evaluated against the test data
• The metric presented below are the ones obtained on the test for the model that generated the best eval metric.

Datasets

Datasets are now sourced from OpenML, usingOpenML:

data_map = Dict(
    :titanic => 40945,
    :higgs_11M => 45570,
    :higgs_1M => 42769,
    :boston => 531,
    :year => 44027,
    :microsoft => 45579,
    :sberbank => 46898, #TODO
    :allstate_claims => 45046, #TODO
    :creditcard => 1597 #TODO
)

Legacy datasets from older release

The following selection of common tabular datasets is covered:

• Year: min squared error regression
• MSRank: ranking problem with min squared error regression
• YahooRank: ranking problem with min squared error regression
• Higgs: 2-level classification with logistic regression
• Boston Housing: min squared error regression
• Titanic: 2-level classification with logistic regression

Algorithms

Comparison is performed against the following algos (implementation in link) considered as state of the art on tabular data problems tasks:

• NeuroTabModels
• EvoTrees
• XGBoost
• LightGBM
• CatBoost
• NODE

Boston

model_type	train_time	test_mse	test_gini
catboost	0.175	0.194	0.945
evotrees	0.198	0.254	0.935
lightgbm	0.314	0.326	0.934
neurotrees	4.58	0.269	0.925
tabm	5.34	0.224	0.934
xgboost	0.0846	0.265	0.93

Titanic

model_type	train_time	test_logloss	test_gini
catboost	0.0759	0.375	0.802
evotrees	0.0399	0.362	0.806
lightgbm	0.209	0.363	0.809
neurotrees	3.15	0.373	0.815
tabm	6.45	0.383	0.774
xgboost	0.0195	0.37	0.795

Year

model_type	train_time	test_mse	test_gini
catboost	65.7	0.621	0.664
evotrees	79.9	0.613	0.666
lightgbm	104.0	0.607	0.67
neurotrees	519.0	0.594	0.68
tabm	27.9	0.616	0.669
xgboost	42.6	0.614	0.666

Microsoft

model_type	train_time	test_mse	test_gini
catboost	186.0	0.73	0.561
evotrees	97.8	0.722	0.567
lightgbm	38.7	0.717	0.571
neurotrees	1110.0	0.76	0.529
tabm	345.0	0.773	0.515
xgboost	42.1	0.719	0.57

Higgs

model_type	train_time	test_logloss	test_gini
catboost	150.0	0.494	0.674
evotrees	55.3	0.496	0.67
lightgbm	67.0	0.495	0.673
neurotrees	291.0	0.487	0.686
tabm	37.6	0.497	0.671
xgboost	35.5	0.496	0.67

References

• Neural Oblivious Decision Ensembles for Deep Learning on Tabular Data.
  • https://github.com/Qwicen/node
  • https://github.com/manujosephv/pytorch_tabular
• Attention augmented differentiable forest for tabular data
• NCART: Neural Classification and Regression Tree for Tabular Data
• Squeeze-and-Excitation Networks
• Deep Neural Decision Trees
• XGBoost: A Scalable Tree Boosting System
• CatBoost: gradient boosting with categorical features support
• LightGBM: A Highly Efficient Gradient Boosting Decision Tree
• Deep Neural Decision Trees
• Neural Decision Trees