Data Distribution-based Curriculum Learning (DDCL)

Copyright © 2025, ECOLS - All rights reserved.

2025-01-30: Initial upload featuring Curriculum Learning and Self-Paced Learning.

1. Static DDCL

Data Distribution-based Curriculum Learning (DDCL) or Static DDCL uses the data distribution of a dataset to build a curriculum based on the order of samples. It includes two types of scoring methods known as DDCL (Density) and DDCL (Point) which determine the difficulty of training samples and thus their order among other samples during training. DDCL (Density) uses the sample density to assign scores while DDCL (Point) utilises the Euclidean distance for scoring.

The Static DDCL approach was developed by Shonal Chaudhry with guidance from Anurag Sharma while being a member of the ECOLS research group.

Paper: S. Chaudhry and A. Sharma, Data Distribution-Based Curriculum Learning, IEEE Access, vol. 12, pp. 138429–138440, 2024, doi: 10.1109/ACCESS.2024.3465793.

2. Dynamic DDCL

Dynamic DDCL extends DDCL by adding a dynamically generated curriculum through self-paced learning. It adds two additional scoring methods known as Self-Paced DDCL (Density) and Self-Paced DDCL (Point) which use a combination of curriculum learning and self-paced learning.

The Dynamic DDCL approach was developed by Shonal Chaudhry with guidance from Anurag Sharma while being a member of the ECOLS research group.

Paper: S. Chaudhry and A. Sharma, Dynamic Data Distribution-based Curriculum Learning, Information Sciences, vol. 702, p. 121924, 2025, doi: (https://doi.org/10.1016/j.ins.2025.121924).

Dependencies

• Python 3.12.7
• Tensorflow 2.17.0
• Pandas 2.2.2
• Scikit-Learn 1.5.1
• Imbalanced-Learn 0.12.3
• Matplotlib 3.9.2
• Seaborn 0.13.2

Running the code

• Run the code through the main.py file. Program behaviour can be altered by changing the options in config.ini.
• There are two sub-program options, DDCL and Plot, which run the DDCL code or generate plots using output files obtained from the DDCL sub-program respectively.

DDCL Examples

Some examples for running a specific DDCL configuration:

Single model - train and predict

1. In config.ini, select one of these learning_model options: Neural Network, SVM or Random Forest.
2. Select a training_strategy option.
3. Use train_and_predict for the nn_task option.

Train and save a model

1. In config.ini, select one of the learning_model options: Neural Network, SVM, Random Forest, DDCL Ensemble or Standard Ensemble.
2. Select a training_strategy option. Note: selecting DDCL Ensemble in Step 1 will use all options.
3. Use train_and_save for the nn_task option.

Predict with ensemble models

1. In config.ini, select either DDCL Ensemble or Standard Ensemble for the learning_model option.
   • For DDCL Ensemble, the training_strategy option is ignored.
   • For Standard Ensemble, select a training_strategy option that has a saved model available.
2. Use load_and_predict for the nn_task option.