A3SOM, abstained explainable semi-supervised neural network based on self-organizing map

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Science

Background:

• Unlabeled data is abundant, posing challenges for supervised learning.
• Label acquisition is often costly and time-consuming.
• Existing methods may fail to identify novel or overlapping classes.

Purpose of the Study:

• To introduce A3SOM, an abstained explainable semi-supervised neural network.
• To enable detection of new classes and class overlaps.
• To provide an explainable model with integrated visualization capabilities.

Main Methods:

• A3SOM integrates a self-organizing map with dense layers for classification.
• Abstained classification is employed to handle uncertain predictions.
• The model incorporates visualization through the self-organizing map.

Main Results:

• A3SOM demonstrates competitive performance against other classifiers.
• The inclusion of abstention rules proves beneficial.
• The method successfully identifies new classes and class overlaps.

Conclusions:

• A3SOM offers a robust and explainable approach to semi-supervised learning.
• The model's abstention mechanism enhances its ability to handle complex datasets.
• A3SOM shows significant potential for real-world applications, such as breast cancer subtype classification.