Explaining a series of models by propagating Shapley values

Area of Science:

• Machine Learning
• Artificial Intelligence
• Explainable AI (XAI)

Background:

• Local feature attribution methods are vital for explaining complex machine learning models.
• Existing methods face limitations due to high computational costs or inability to handle distributed model series.
• Distributed model explanations are critical in regulated industries like finance.

Purpose of the Study:

• To introduce Generalized DeepSHAP (G-DeepSHAP), a computationally tractable method for propagating local feature attributions.
• To address the limitations of current attribution techniques in distributed and complex model settings.
• To provide efficient and salient explanations for machine learning models in various domains.

Main Methods:

• Developed Generalized DeepSHAP (G-DeepSHAP) based on Shapley value principles.
• Propagated local feature attributions through complex, distributed model series.
• Evaluated G-DeepSHAP on biological, health, and financial datasets.

Main Results:

• G-DeepSHAP provides equally salient explanations compared to existing methods.
• G-DeepSHAP is an order of magnitude faster than current model-agnostic attribution techniques.
• Demonstrated the utility of G-DeepSHAP in a distributed series of models scenario.

Conclusions:

• G-DeepSHAP is an efficient and effective solution for local feature attribution in complex and distributed machine learning models.
• The method significantly improves computational efficiency without sacrificing explanation quality.
• G-DeepSHAP has practical implications for industries requiring model interpretability, such as finance.