A Unifying Post-Processing Framework for Multi-Objective Learn-to-Defer Problems

Authors: Mohammad-Amin Charusaie, Samira Samadi

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

Reproducibility Variable Result LLM Response
Research Type Experimental Furthermore, we design a generalizable algorithm to estimate that solution and apply this algorithm to the COMPAS, Hatespeech, and ACSIncome datasets. Our algorithm shows improvements in terms of constraint violation over a set of learn-to-defer baselines and can control multiple constraint violations at once.
Researcher Affiliation Academia Mohammad-Amin Charusaie Max Planck Institute for Intelligent Systems Tuebingen, Germany mcharusaie@tuebingen.mpg.de Samira Samadi Max Planck Institute for Intelligent Systems Tuebingen, Germany samira.samadi@tuebingen.mpg.de
Pseudocode Yes Based on this optimal solution, we can design a plug-in method (see Algorithm 1 in Appendix F) to solve the constrained learning problem using empirical data.
Open Source Code Yes The code is available in https://github.com/Amin Chrs/Post Process/.
Open Datasets Yes Our algorithm shows improvements in terms of constraint violation over a set of learn-to-defer baselines and can control multiple constraint violations at once. The use of d-GNP is beyond learn-to-defer applications and can potentially obtain a solution to decision-making problems with a set of controlled expected performance measures.
Dataset Splits Yes n is the size of the set using which we fine-tune the algorithm, ϵ measures the accuracy of learned post-processing scores, and γ is a parameter that measures the sensitivity of the constraint to the change of the predictor.
Hardware Specification No The paper does not provide specific details about the hardware (e.g., GPU/CPU models, memory) used for running the experiments.
Software Dependencies No The paper mentions training on a '1-layer feed-forward neural network' and using a 'pre-trained model [5]' but does not specify any software libraries or frameworks with version numbers (e.g., PyTorch, TensorFlow, Python version).
Experiment Setup Yes All scores, classifiers, and rejection functions are trained on a 1-layer feed-forward neural network. The human assessment is done in this dataset on 1000 cases by giving humans a description of the case and asking them whether the defendant would recidivate within two years of their most recent crime.