Towards Robust and Reliable Algorithmic Recourse

Authors: Sohini Upadhyay, Shalmali Joshi, Himabindu Lakkaraju

NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental evaluation on multiple synthetic and real-world datasets demonstrates the efficacy of the proposed framework. We evaluated our approach ROAR on real world data from financial lending and education domains...
Researcher Affiliation Academia Sohini Upadhyay Harvard University supadhyay@g.harvard.edu Shalmali Joshi Harvard University shalmali@seas.harvard.edu Himabindu Lakkaraju Harvard University hlakkaraju@hbs.harvard.edu
Pseudocode Yes Algorithm 1 Our Optimization Procedure
Open Source Code No The paper does not provide concrete access to source code (specific repository link, explicit code release statement, or code in supplementary materials) for the methodology described in this paper.
Open Datasets Yes Our first dataset is the widely used and publicly available German credit dataset [8] from the UCI repository. Our second dataset is the Small Business Administration (SBA) case dataset [17]. Our last dataset contains student performance records of 649 students from two Portuguese secondary schools, Gabriel Pereira (GP) and Mousinho da Silveira (MS) [8, 6].
Dataset Splits Yes We use 5-fold cross validation throughout our real world and synthetic experiments. On D1, we use 4 folds to train predictive models and the remaining fold to generate and evaluate recourses.
Hardware Specification No The paper does not provide specific hardware details (exact GPU/CPU models, processor types with speeds, memory amounts, or detailed computer specifications) used for running its experiments.
Software Dependencies No The paper only mentions 'binary cross entropy loss and the Adam optimizer' without providing specific version numbers for these or any other software dependencies.
Experiment Setup Yes Our framework, ROAR, has the following parameters: the set of acceptable perturbations (defined in practice by δmax) and the tradeoff parameter λ. In our experiments on evaluating robustness to real world shifts, we choose δmax = 0.1 given that continuous features are scaled to zero mean and unit variance. We use binary cross entropy loss and the Adam optimizer to operationalize our framework, ROAR.