Explainable Data-Driven Optimization: From Context to Decision and Back Again

Authors: Alexandre Forel, Axel Parmentier, Thibaut Vidal

ICML 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We demonstrate the value of our methods to explain data-driven decisions in repeated experiments with synthetic and real-world data.
Researcher Affiliation Academia 1CIRRELT & SCALE-AI Chair in Data-Driven Supply Chaims, Department of Mathematical and Industrial Engineering, Polytechnique Montreal, Montreal, Canada 2CERMICS, École des Ponts, Marne-la-Vallée, France.
Pseudocode Yes Algorithm 1 Iterative procedure for absolute explanations
Open Source Code Yes The code used to generate all the results in this paper is publicly available at https://github. com/alexforel/Explainable-CSO under an MIT license.
Open Datasets Yes Multi-item newsvendor. This problem is adapted from Kallus & Mao (2022). Two-stage shipment planning. This problem is taken from Bertsimas & Kallus (2020). (CVa R) shortest path. The problem is adapted from Elmachtoub et al. (2020) and Elmachtoub & Grigas (2022). Historical observations of the uncertain travel times are available on the edge level. We randomly sample n = 1000 past observations and train a random forest predictor.
Dataset Splits No The paper mentions 'The training data {(xi, yi)}n i=1 [...] is resampled in each experiment' and 'We randomly sample n = 1000 past observations and train a random forest predictor', but it does not specify explicit training/validation/test splits or a dedicated validation methodology.
Hardware Specification Yes All experiments are run on four cores of an Intel(R) Xeon(R) Gold 6336Y CPU at 2.40 GHz.
Software Dependencies Yes The simulations are implemented in Python 3.9.13. Gurobi 9.5.1 is used to solve all mixed-integer programming problems. [...] We use the standard procedure of scikit-learn v.1.0.2.
Experiment Setup Yes We use the standard procedure of scikit-learn v.1.0.2. with T = 100 trees (default value) and a maximum depth of 4. We focus on nearest-neighbor predictors with k = 10. All distances are measured through the l1 norm...