Use-Case-Grounded Simulations for Explanation Evaluation

Authors: Valerie Chen, Nari Johnson, Nicholay Topin, Gregory Plumb, Ameet Talwalkar

NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We run a comprehensive evaluation on three real-world use cases (forward simulation, model debugging, and counterfactual reasoning) to demonstrate that Sim Evals can effectively identify which explanation methods will help humans for each use case.
Researcher Affiliation Academia Valerie Chen Nari Johnson Nicholay Topin Gregory Plumb Ameet Talwalkar Carnegie Mellon University Correspondence to valeriechen@cmu.edu.
Pseudocode No The paper does not contain any pseudocode or algorithm blocks.
Open Source Code Yes Tutorial on how to run your own Sim Evals: https://github.com/valeriechen/simeval_tutorial
Open Datasets Yes A base dataset (D = {(x, y)}), on which the explanation method s utility will be evaluated. For example, the data generation process for the forward simulation use case from Hase and Bansal [16]. (Left) For this particular information content type, each observation contains the data-point xi and the model explanation for that datapoint E(f, xi). The three use-case-specific components are the base dataset D, base model family F, and function that defines a use case label. (Right) These components are used to generate a dataset of N observations. (x1 , E( x1 , f )), f(x1) (x N , E( x N , f )), f(x N ) (x2 , E( x2 , f )), f(x2) xi , yi From : UCI Adult
Dataset Splits Yes Following the same procedure as Hase and Bansal, we split D into train, test, and validation sets.
Hardware Specification No The paper does not specify any particular hardware (e.g., CPU, GPU models, or cloud instance types) used for running the experiments.
Software Dependencies No The paper mentions various explanation methods and models like LIME, SHAP, GAM, LightGBM, but it does not provide specific version numbers for any software dependencies needed to reproduce the experiments.
Experiment Setup Yes We include training details for our Sim Evals in Appendix F, extended results (with error bars) and comparisons with human results in Appendix G.