Skip to main content

transfer parameters from lightgbm to differentiable decision trees!

Project description


TreeGrad implements a naive approach to converting a Gradient Boosted Tree Model to an Online trainable model. It does this by creating differentiable tree models which can be learned via auto-differentiable frameworks. TreeGrad is in essence an implementation of Kontschieder, Peter, et al. "Deep neural decision forests." with extensions.

To install

python install

or alternatively from pypi

pip install treegrad

Run tests:

python -m nose2
  title={Transferring Tree Ensembles to Neural Networks},
  author={Siu, Chapman},
  booktitle={International Conference on Neural Information Processing},



from sklearn.
import treegrad as tgd

mod = tgd.TGDClassifier(num_leaves=31, max_depth=-1, learning_rate=0.1, n_estimators=100, autograd_config={'refit_splits':False}), y)
mod.partial_fit(X, y)


The requirements for this package are:

  • lightgbm
  • scikit-learn
  • autograd

Future plans:

  • Add implementation for Neural Architecture search for decision boundary splits (requires a bit of clean up - TBA)
    • Implementation can be done quite trivially using objects residing in - Challenge is getting this working in a sane manner with scikit-learn interface.
  • GPU enabled auto differentiation framework - see notebooks/ for progress off Colab for Tensorflow 2.0 port
  • support xgboost/lightgbm additional features such as monotone constraints
  • Support RegressorMixin


When decision splits are reset and subsequently re-learned, TreeGrad can be competitive in performance with popular implementations (albeit an order of magnitude slower). Below is a table showing accuracy on test dataset on UCI benchmark datasets for Boosted Ensemble models (100 trees)

Dataset TreeGrad LightGBM Scikit-Learn (Gradient Boosting Classifier)
adult 0.860 0.873 0.874
covtype 0.832 0.835 0.826
dna 0.950 0.949 0.946
glass 0.766 0.813 0.719
mandelon 0.882 0.881 0.866
soybean 0.936 0.936 0.917
yeast 0.591 0.573 0.542


To understand the implementation of TreeGrad, we interpret a decision tree algorithm to be a three layer neural network, where the layers are as follows:

  1. Node layer, which determines the decision boundaries
  2. Routing layer, which determines which nodes are used to route to the final leaf nodes
  3. Leaf layer, the layer which determines the final predictions

In the node layer, the decision boundaries can be interpreted as axis-parallel decision boundaries from your typical Linear Classifier; i.e. a fully connected dense layer

The routing layer requires a binary routing matrix to which essentially the global product routing is applied

The leaf layer is your typical fully connected dense layer.

This approach is the same as the one taken by Kontschieder, Peter, et al. "Deep neural decision forests."

Project details

Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Files for treegrad, version 1.0.1
Filename, size File type Python version Upload date Hashes
Filename, size treegrad-1.0.1-py3-none-any.whl (11.7 kB) File type Wheel Python version py3 Upload date Hashes View

Supported by

Pingdom Pingdom Monitoring Google Google Object Storage and Download Analytics Sentry Sentry Error logging AWS AWS Cloud computing DataDog DataDog Monitoring Fastly Fastly CDN DigiCert DigiCert EV certificate StatusPage StatusPage Status page