Approximating the Shapley Value without Marginal Contributions

Authors: Patrick Kolpaczki, Viktor Bengs, Maximilian Muschalik, Eyke Hüllermeier

AAAI 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We prove unmatched theoretical guarantees regarding their approximation quality and provide empirical results including synthetic games as well as common explainability use cases comparing ourselves with state-of-the-art methods.
Researcher Affiliation Academia 1Paderborn University 2Institute of Informatics, University of Munich (LMU) 3Munich Center for Machine Learning patrick.kolpaczki@upb.de, viktor.bengs@lmu.de, maximilian.muschalik@lmu.de, eyke@lmu.de
Pseudocode Yes Algorithm 1: SVARM, Algorithm 2: Stratified SVARM
Open Source Code Yes All code including documentation and the technical appendix can be found on Git Hub1. 1https://github.com//kolpaczki//Approximating-the-Shapley Value-without-Marginal-Contributions
Open Datasets Yes NLP sentiment analysis game is based on the Distil BERT (Sanh et al. 2019) model architecture and consists of randomly selected movie reviews from the IMDB dataset (Maas et al. 2011) containing 14 words. In the image classifier game, we explain the output of a Res Net18 (He et al. 2016) trained on Image Net (Deng et al. 2009). For the adult classification game, we train a gradient-boosted tree model on the adult dataset (Becker and Kohavi 1996).
Dataset Splits No The paper mentions using specific datasets (IMDB, ImageNet, Adult) for the explainability tasks but does not specify the train/validation/test splits used for the training of the underlying machine learning models that are then explained. The focus is on approximating Shapley values for already existing or exhaustively calculated models, not on the data splits for training those models.
Hardware Specification No The paper does not provide specific details regarding the hardware (e.g., GPU/CPU models, memory, cloud resources) used for conducting the experiments.
Software Dependencies No The paper does not list specific software dependencies with their version numbers required to reproduce the experiments. While it mentions models like Distil BERT and Res Net18, the underlying software frameworks and their versions are not specified.
Experiment Setup No The paper describes the general experimental setup for synthetic and explainability games, including details on how coalitions are valued (e.g., masking, mean imputation, super-pixels). It also specifies the performance metric (MSE) and evaluation over a range of fixed budget values T. However, it does not provide specific hyperparameters or system-level training configurations (e.g., learning rates, batch sizes, specific optimizer settings, or model initialization details) for the algorithms presented or the underlying ML models.