A Regularization Approach for Instance-Based Superset Label Learning

Area of Science:

• Machine Learning
• Artificial Intelligence
• Computer Science

Background:

• Superset label learning (SLL) involves training examples with multiple candidate labels, only one of which is correct.
• Existing instance-based SLL methods achieve high performance but struggle with non-discriminative label disambiguation due to ignoring label exclusivity.

Purpose of the Study:

• To develop a novel regularization approach for instance-based superset label learning (RegISL).
• To enhance the discriminative ability of instance-based SLL methods by incorporating regularization techniques.

Main Methods:

• A graph is employed to represent the training set, enforcing similar labels for adjacent examples.
• A novel discrimination term is introduced to increase the distinction between likely and unlikely labels for each training example.

Main Results:

• RegISL generates more discriminative and accurate disambiguated labels by leveraging intrinsic constraints among candidate labels.
• Experimental results demonstrate RegISL's superiority over existing SLL methods in both training and test accuracy.

Conclusions:

• The proposed RegISL method effectively addresses the limitations of current instance-based SLL approaches.
• RegISL enhances label disambiguation accuracy and overall performance in superset label learning tasks.