Incremental Class Learning for Hierarchical Classification.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Science

Background:

• Objects are perceived and described hierarchically, necessitating hierarchical classification in machine learning.
• Existing classification methods require complete retraining for new data classes, hindering incremental learning.

Purpose of the Study:

• To propose an adaptive resonance theory-supervised predictive mapping for hierarchical classification (ARTMAP-HC) network.
• To enable incremental class learning for raw data without prior normalization.
• To address the limitations of retraining in hierarchical classification tasks.

Main Methods:

• ARTMAP-HC utilizes hierarchically stacked modules, each containing two fuzzy ARTMAP networks.
• Incorporates a novel online normalization process for data classification without prior knowledge of dataset maximums.
• Employs a prior labels appending process to reflect class dependencies across hierarchy levels.

Main Results:

• ARTMAP-HC demonstrates effective incremental learning for sequentially added input data, regardless of hierarchy level or class count.
• The network successfully classifies new data without prior normalization, showcasing its adaptive capabilities.
• Experiments on hierarchical classification datasets validate the effectiveness of the proposed ARTMAP-HC approach.

Conclusions:

• ARTMAP-HC provides an efficient solution for hierarchical classification with incremental learning capabilities.
• The network's ability to learn without retraining and handle raw data makes it suitable for dynamic datasets.
• ARTMAP-HC shows promise for applications like multimedia recommendation systems, enhancing digital storytelling.