SHAPでの説明をするツールを紹介したい。
シュープレイ値について(Wikipedia)
https://en.wikipedia.org/wiki/Shapley_value#In_machine_learning
・SHAP (SHapley Additive exPlanations) DeepSHAP
https://github.com/slundberg/shap
このツールは
・Tree ensemble example (XGBoost/LightGBM/CatBoost/scikit-learn/pyspark models)
・Natural language example (transformers)
・Deep learning example with DeepExplainer (TensorFlow/Keras models)
・Deep learning example with GradientExplainer (TensorFlow/Keras/PyTorch models)
・Model agnostic example with KernelExplainer (explains any function)
・SHAP Interaction Values
wikipediaでの参考文献
・SHAP値のmachine learningへの導入のきっかけとなった論文
Lundberg, Scott M.; Lee, Su-In. “A Unified Approach to Interpreting Model Predictions”. Advances in Neural Information Processing Systems. 30: 4765–4774. Retrieved 2021-01-30.
https://papers.nips.cc/paper/2017/hash/8a20a8621978632d76c43dfd28b67767-Abstract.html
・Locally Interpretable Model-Agnostic Explanations (LIME)
Ribeiro, Marco Tulio; Singh, Sameer; Guestrin, Carlos (2016-08-13). “Why Should I Trust You?”. New York, NY, USA: ACM.
・DeepLIFT
https://github.com/kundajelab/deeplift
Shrikumar, Avanti; Greenside, Peyton; Kundaje, Anshul (2017-07-17). “Learning Important Features Through Propagating Activation Differences”. PMLR. pp. 3145–3153. ISSN 2640-3498. Retrieved 2021-01-30.
・Layer-Wise Relevance Propagation
Bach, Sebastian; Binder, Alexander; Montavon, Grégoire; Klauschen, Frederick; Müller, Klaus-Robert; Samek, Wojciech (2015-07-10). Suarez, Oscar Deniz (ed.). “On Pixel-Wise Explanations for Non-Linear Classifier Decisions by Layer-Wise Relevance Propagation”. PLOS ONE. Public Library of Science (PLoS). 10 (7): e0130140. doi:10.1371/journal.pone.0130140. ISSN 1932-6203.
・DeepExplain
https://github.com/marcoancona/DeepExplain
経緯としてはICML2018でAnconaらのDeepExplainを契機にして、DeepLIFTとDeepSHAPが生まれたらしい。
どちらかの違いについてはDeepLIFTのFAQに詳しい。結論としてはまずDeepSHAPでやり、データの規模が大きい場合はDeepLIFTがおすすめらしい。
SHAPのメソッドを試した記事
・https://qiita.com/shin_mura/items/cde01198552eda9146b7
・https://dropout009.hatenablog.com/entry/2019/11/20/091450?_ga=2.185612329.1656579419.1574324138-1185090719.1545096259
