ILLC: Iterative Layer-by-Layer Compression for Enhancing Structural Faithfulness in SpArX

Layer-by-layer compression preserves model logic better than traditional methods: The paper demonstrates that compressing neural networks one layer at a time, rather than all at once, reduces information loss and maintains more accurate representations of how the model actually makes decisions—important for trustworthy AI systems.

Argumentative explanations can be distorted during compression: The research shows that existing compression techniques can damage the "attack-support relationships" in argumentative explanations (how the model's internal reasoning chains together), meaning simplified models may misrepresent the actual logic behind AI decisions.

The iterative approach immediately corrects errors as they accumulate: By compensating for compression errors in each layer before moving to the next, this method prevents the "accumulative information loss" that occurs when all layers are simplified simultaneously, leading to more faithful explanations of model behavior.

Structural fidelity matters alongside accuracy: The paper emphasizes that it's not enough for compressed models to produce correct outputs—they must also preserve the internal structure and reasoning process, which is critical for users to understand and trust how AI systems work.

Tested on real medical AI applications: Validation on a Breast Cancer Diagnosis dataset demonstrates practical relevance for high-stakes domains where both model accuracy and explainability are essential for responsible deployment.