Dynamic sampling approach to training neural networks for multiclass imbalance classification.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Science

Background:

• Class imbalance is a common challenge in supervised learning where datasets have uneven class distributions.
• Existing research predominantly addresses binary-class imbalance, leaving multiclass scenarios less explored.
• Multilayer perceptrons (MLPs) are widely used but can struggle with imbalanced multiclass data.

Purpose of the Study:

• To propose and evaluate a novel dynamic sampling method (DyS) for addressing multiclass imbalance problems in MLPs.
• To investigate the effectiveness of DyS in selecting informative data points during the training process.
• To compare the performance of DyS against various established methods for imbalanced learning.

Main Methods:

• A dynamic sampling (DyS) approach is proposed for training MLPs on multiclass imbalanced datasets.
• DyS estimates the probability of selecting each example for training in each epoch.
• The method dynamically focuses on informative data to improve model training.

Main Results:

• Comprehensive experiments were conducted on 20 multiclass imbalanced datasets.
• DyS demonstrated superior performance compared to pre-sampling, active learning, cost-sensitive, and boosting methods.
• The study highlights the strengths and weaknesses of the DyS approach through empirical evaluation.

Conclusions:

• The proposed dynamic sampling (DyS) method is effective for tackling multiclass imbalanced learning problems with MLPs.
• DyS offers a promising alternative to existing methods by adaptively selecting training data.
• Further research can explore DyS in different neural network architectures and application domains.