EvanZhuang / MetaTree

Official implementation of MetaTree: Learning a Decision Tree Algorithm with Transformers
https://arxiv.org/abs/2402.03774
MIT License
101 stars 5 forks source link

🌲 MetaTree 🌲

Learning a Decision Tree Algorithm with Transformers (Zhuang et al., TMLR 2024).

MetaTree is a transformer-based decision tree algorithm. It learns from classical decision tree algorithms (greedy algorithm CART, optimal algorithm GOSDT), for better generalization capabilities.

Quickstart -- use MetaTree to generate decision tree models

Model is available at https://huggingface.co/yzhuang/MetaTree

  1. Install metatreelib:
pip install metatreelib
# Alternatively,  
# clone then pip install -e .
# pip install git+https://github.com/EvanZhuang/MetaTree
  1. Use MetaTree on your datasets to generate a decision tree model
from metatree.model_metatree import LlamaForMetaTree as MetaTree
from metatree.decision_tree_class import DecisionTree, DecisionTreeForest
from metatree.run_train import preprocess_dimension_patch
from transformers import AutoConfig
import imodels # pip install imodels 

# Initialize Model
model_name_or_path = "yzhuang/MetaTree"

config = AutoConfig.from_pretrained(model_name_or_path)
model = MetaTree.from_pretrained(
    model_name_or_path,
    config=config,
)
decision_tree_forest = DecisionTreeForest()   

# Load Datasets
X, y, feature_names = imodels.get_clean_dataset('fico', data_source='imodels')

print("Dataset Shapes X={}, y={}, Num of Classes={}".format(X.shape, y.shape, len(set(y))))

train_idx, test_idx = sklearn.model_selection.train_test_split(range(X.shape[0]), test_size=0.3, random_state=seed)

# Dimension Subsampling
feature_idx = np.random.choice(X.shape[1], 10, replace=False)
X = X[:, feature_idx]

test_X, test_y = X[test_idx], y[test_idx]

# Sample Train and Test Data
subset_idx = random.sample(train_idx, 256)
train_X, train_y = X[subset_idx], y[subset_idx]

input_x = torch.tensor(train_X, dtype=torch.float32)
input_y = torch.nn.functional.one_hot(torch.tensor(train_y)).float()

batch = {"input_x": input_x, "input_y": input_y, "input_y_clean": input_y}
batch = preprocess_dimension_patch(batch, n_feature=10, n_class=10)
model.depth = 2
outputs = model.generate_decision_tree(batch['input_x'], batch['input_y'], depth=model.depth)
decision_tree_forest.add_tree(DecisionTree(auto_dims=outputs.metatree_dimensions, auto_thresholds=outputs.tentative_splits, input_x=batch['input_x'], input_y=batch['input_y'], depth=model.depth))

print("Decision Tree Features: ", [x.argmax(dim=-1) for x in outputs.metatree_dimensions])
print("Decision Tree Thresholds: ", outputs.tentative_splits)
  1. Inference with the decision tree model
tree_pred = decision_tree_forest.predict(torch.tensor(test_X, dtype=torch.float32))

accuracy = accuracy_score(test_y, tree_pred.argmax(dim=-1).squeeze(0))
print("MetaTree Test Accuracy: ", accuracy)

Example Usage

We show a complete example of using MetaTree at notebook

Questions?

If you have any questions related to the code or the paper, feel free to reach out to us at y5zhuang@ucsd.edu.

Citation

If you find our paper and code useful, please cite us:

@misc{zhuang2024learning,
      title={Learning a Decision Tree Algorithm with Transformers}, 
      author={Yufan Zhuang and Liyuan Liu and Chandan Singh and Jingbo Shang and Jianfeng Gao},
      year={2024},
      eprint={2402.03774},
      archivePrefix={arXiv},
      primaryClass={cs.LG}
}