Learning Concept Credible Models for Mitigating Shortcuts.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Science

Background:

• Models often learn spurious correlations (shortcuts) during training, leading to poor generalization on new data.
• Existing methods struggle when shortcuts are linked to unknown concepts or when relying solely on predefined knowledge.

Purpose of the Study:

• To develop methods for learning robust and accurate models from biased training data by mitigating shortcut learning.
• To incorporate domain knowledge (known concepts) while also allowing models to learn unknown concepts.

Main Methods:

• A two-stage approach: first fitting a model with known concepts, then addressing residuals with unknown concepts.
• An extended regularization penalty approach that integrates known and unknown concepts to prevent shortcut exploitation.

Main Results:

• The two-stage approach shows vulnerability when shortcuts correlate with unknown concepts.
• The extended regularization penalty approach effectively mitigates shortcut learning across two real-world datasets.

Conclusions:

• Both proposed methods successfully reduce shortcut learning in models trained on biased data.
• Integrating domain knowledge with the ability to learn unknown concepts is crucial for robust model generalization.