BONES: a Benchmark fOr Neural Estimation of Shapley values

Shapley Values help explain AI decisions but are computationally expensive: The paper addresses a real problem—while Shapley Values are useful for understanding why AI models make specific predictions, calculating them exactly on real-world datasets is impractical, making faster approximations necessary.

Neural networks can estimate Shapley Values more quickly: The research demonstrates that machine learning-based estimators can approximate Shapley Values much faster than traditional exact methods, potentially making explainability more feasible for real-world AI systems.

There was no standardized way to compare different explanation methods: Before this work, researchers lacked common tools and datasets to fairly evaluate and compare different approaches to estimating Shapley Values, making it difficult to know which methods worked best—BONES provides this standardized benchmark.

The benchmark works across different types of data: BONES is tested on both tabular data (spreadsheet-like information) and image data, suggesting the tool could be useful for explaining AI decisions across many different applications and industries.

Open-source tool aims to democratize AI explainability research: By releasing BONES as free, open-source software, the authors are making it easier for researchers and practitioners to implement, test, and compare explanation methods without building tools from scratch.