/Deep-reinforcement-learning-for-multi-class-imbalanced-classification

Implementation of "Deep reinforcement learning for imbalanced classification" and its extended version to multi-class

Primary LanguagePythonMIT LicenseMIT

Deep-reinforcement-learning-for-multi-class-imbalanced-classification

Implementation of Deep reinforcement learning for imbalanced classification and its extended version to multi-class imbalanced classification.

Differences with the original paper

  • Double DQN and Dueling DQN are applied.
  • The reward function on the paper is extended to multi-class imbalanced data.
  • It has been implemented to easily test various multi-class imbalanced settings of Cifar-10 dataset.

Test environment

  • Python 3.7
  • Tensorflow 1.14

Examples of how to run

You can check example codes for some major configurations in demo.sh.

$ ./demo.sh

Experiment results

The values of train parameters from the original paper are used.

Dataset Imbalance ratio F-measure
Cifar-10(1) 4% 0.901
2% 0.879
1% 0.862
0.5% 0.784
Cifar-10(2) 4% 0.887
2% 0.855
1% 0.806
0.5% 0.708