DIFER

Code for "DIFER: Differentiable Automated Feature Engineering"

accepted in 1st Conference on Automated Machine Learning

Requisite

This code is implemented in PyTorch, and we have tested the code under the environment settings in requirements.txt.

What is in this repository

• data: $23/25$ medium-sized datasets that can be pushed to git and their meta information.
• NFS_sklearn_c: the open-source implementation of "Neural Feature Search: A Neural Architecture for Automated Feature Engineering".
• autolearn:the core coes for DIFER in autolearn/feat_selection/nfo, continas the feature optimizer in controller.py, the feature space in search_space.py, the end-to-end training process in iter_train.py, the three forms of feature (i.e., the original form, the parse tree and the traversal string) in feat_tree.py.

How to run it

We provide script files for convenience in conducting experiments.

• run_iter.sh: after specifying the dataset and cuda, you can run DIFER to automate feature engineering for Random Forest.
• run_rq3.sh: the script for RQ3 in the paper.
• run_rq4_*.sh: the script of different machine learning algorithms for RQ4 in the paper.

Reference Code

• NFS: https://github.com/TjuJianyu/NFS