Soft Hoeffding Tree

This repository contains a new predictive base learner Soft Hoeffding Tree (SoHoT) for large and drifting data streams. A SoHoT combines the extensibility and transparency of Hoeffding trees with the differentiability of soft trees. A new gating function is used to regulate the balance between transparent and soft routing in a SoHoT, which is essentially a trade-off between transparency and performance. SoHoT can be used in a neural network due to its differentiability.

Example Code

from benchmark.load_data import load_data_stream
from sohot.sohot_ensemble_layer import SoftHoeffdingTreeLayer
from capymoa.evaluation import ClassificationEvaluator


data_stream, n_instances = load_data_stream('AGR_small')
model = SoftHoeffdingTreeLayer(schema=data_stream.get_schema(), trees_num=1)
evaluator = ClassificationEvaluator(schema=data_stream.get_schema())
for i in range(n_instances):
    instance = data_stream.next_instance()
    y_pred = model.predict(instance)
    evaluator.update(y_target_index=instance.y_index, y_pred_index=y_pred)
    model.train(instance)
    if i == 5000:
        model.plot_tree(instance=instance, tree_idx=0)
print(f"Accuracy: {evaluator.accuracy()}")

Transparency

Soft Hoeffding tree evolving over the data stream generated with Agrawal generator containing three abrupt drifts (applying the classification functions: 2, 3, 5, 6).

Cite SoHoT

First introduction of SoHoTs:

@InProceedings{10.1007/978-3-031-78977-9_11,
    author      =   "K{\"o}bschall, Kirsten and Hartung, Lisa and Kramer, Stefan",
    title       =   "Soft Hoeffding Tree: A Transparent and Differentiable Model on Data Streams",
    booktitle   =   "Discovery Science",
    year        =   "2025",
    publisher   =   "Springer Nature Switzerland",
    address     =   "Cham",
    pages       =   "167--182",
    isbn        =   "978-3-031-78977-9"
}

Journal article:

@article{10.1007/s10994-025-06906-x,
  title         =   "Adaptive differentiable trees for transparent learning on data streams",
  author        =   "K{\"o}bschall, Kirsten and Hartung, Lisa and Kramer, Stefan",
  journal       =   "Machine Learning",
  publisher     =   "Springer Science and Business Media LLC",
  doi           =   "10.1007/s10994-025-06906-x",
  volume        =   114,
  number        =   11,
  month         =   nov,
  year          =   2025
}