A Comparative Survey: Benchmarking for Pool-based Active Learning

This is an implementation of our IJCAI 2021 Survey Track paper "A Comparative Survey: Benchmarking for Pool-based Active Learning"

---

Code

We utilized three existing libraries here: libact, google active learning toolbox and ALiPy.

• libact: US, QBC, HintSVM, QUIRE, VR, ALBL, DWUS
• Google active learning: Random/Uniform sampling, KCenter, Graph, Margin, Hier, MCM
• ALiPy: EER, LAL, BMDR, SPAL

---

Prerequisites

Major libraries are listed here:

• libact
• ALiPy
• scikit-learn
• cvxpy
• numpy
• scipy
• tensorflow

You can also use the following command to install conda environment

conda env create -f environment.yml

---

Data

We employed 35 public datasets, which are uniformly dealed with libsvm style, as shown in Dataset/dealeddata/ file folder.

---

Demo

Here we take appendicitis dataset as binary-class classification task example and thyroid dataset as multi-class classification task example, with batch size 5, repeat 100 trials, initial labeled set with size 20.

General format: python baseline-[library name]-[binary/mulitple].py [dataset name] [AL model name] [AL batch size] [number of repeat trials] [initial labeled set size]

• libact: take AL algorithm QBC
  1. binary-class classification: python baseline-libact-binary.py appendicitis-svmstyle QBC 5 100 20
  2. multi-class classification: python baseline-libact-multiple.py thyroid-svmstyle QBC 5 100 20
• Google active learning: take AL algorithm Hier as example
  1. binary-class classification: python baseline-google-binary.py appendicitis-svmstyle Hier 5 100 20
  2. multi-class classification: python baseline-google-multiple.py thyroid-svmstyle Hier 5 100 20
• ALiPy: take AL algorithm EER as example
  1. binary-class classification: python baseline-alipy-binary.py appendicitis-svmstyle QueryExpectedErrorReduction 5 100 20
  2. multi-class classification: python baseline-alipy-multiple.py thyroid-svmstyle QueryExpectedErrorReduction 5 100 20

Please read notes.txt before running the code, especially for BMDR, SPAL and LAL!!!

---

Citing

If you use our code in your research or applications, please consider citing our paper.

@inproceedings{zhan2021comparative,
  title={A Comparative Survey: Benchmarking for Pool-based Active Learning.},
  author={Zhan, Xueying and Liu, Huan and Li, Qing and Chan, Antoni B},
  booktitle={IJCAI},
  pages={4679--4686},
  year={2021}
}

---

Contact

If you have any further questions or want to discuss Active Learning with me, please contact xyzhan2-c@my.cityu.edu.hk (my spare email is sinezhan17@gmail.com).