KO | EN

GitLite

Minimalist Open Source Repository Search Engine

gitlite — search

#typescript #ai-agents #developer-tools #claude-code #codex #cli #agent-skills #ai #dsh-plugin #mcp #windows #agent

Federated-XGBoost

★ 46 Open GitHub ↗

Federated Learning on XGBoost

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

# Federated-XGBoost
Federated Learning on XGBoost
# 
This repository is the implementation of the paper " [The Tradeoff Between Privacy and Accuracy in Anomaly Detection Using Federated XGBoost](https://arxiv.org/abs/1907.07157) ".

## Dataset

Dataset used in this experiment is from kaggle : [Credit Card Fraud Detection](https://www.kaggle.com/mlg-ulb/creditcardfraud).

The aggregated data in the process of data aggregation has been put into the Data folder. Inside each aggregated data folder, there are three files.   `creditcard1_train` and `creditcard1_test` is used for training and in server node. `creditcard2` and  `credit2_for_update` are the dataset in local node.
 `credit2_for_update` is the data `creditcard2` after the process of federated update. The index of each aggregated data folder is the cluster number we want to set in process of finding split.

## Run

We are implmenting the horizontal federated learning scenario based on [XGBoost](https://github.com/dmlc/xgboost).

Firstly, download the XGBoost package following the [XGBoost official documentation](https://xgboost.readthedocs.io/en/latest/).  In order to achieve the federated framework of our paper, there are two files that need to be modified. File `param.h` and `updater_histmaker.cc` have been put into folder Code. Exact location is in `xgboost/python-package/xgboost/src/tree`.

Main change is to set a new parameter for our aggregated sequence in file `param.h` (Line 99 and 103) and add the corresponding way of creating split set in the paper (changes are in `updater_histmaker.cc` Line 292, 313 and 493). Remember to recompile it after adding changes.

Some other steps may also need to be done before running.  Comment Line 11、53 and 54 in file `submit.py` because there are some errors when running code. Also, we need XGBoost rabit to achieve communications between nodes. So, `cd  /xgboost/rabit` and do `make`.

We add parameter `fl_split` in federated XGBoost, which is used to set the cluster number for training. For example, in our file `data_map405`, we map the original instances values into new sequence with 405 clusters. So you can set up that parameter for our aggregated dataset. Like

> XGBClassifier( fl_split=405, tree_method='approx'.....)

Federated model is produced by 
> /xgboost/dmlc-core/tracker/dmlc-submit --cluster mpi --num-workers 2 python main.py

Since the dataset is small, we run it locally and set  two nodes (num-workers =2) to simulate the experiments. You can change  parameters according to your enviroment.
## 
For now, if you want to try it on new dataset, you have to do the process of data aggregation and federated update. Over the next, we would try to experiment it on large dataset and explore some new methods to improve these two processes.

Welcome to contact us if you have any quastion about the paper or code.

## Reference

>@article{yang2019tradeoff,
  title={The Tradeoff Between Privacy and Accuracy in Anomaly Detection Using Federated XGBoost},
  author={Yang, Mengwei and Song, Linqi and Xu, Jie and Li, Congduan and Tan, Guozhen},
  journal={arXiv preprint arXiv:1907.07157},
  year={2019}
}

Federated-XGBoost

Explore Similar Repositories

Was this content helpful?

Recent Feedback

Download README