FasterRisk: Fast and Accurate Interpretable Risk Scores
Authors: Jiachang Liu, Chudi Zhong, Boxuan Li, Margo Seltzer, Cynthia Rudin
NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Through extensive experiments, we show that our proposed method is computationally fast and produces high-quality integer solutions. We experimentally focus on two questions: (1) How good is Faster Risk s solution quality compared to baselines? ( 4.1) (2) How fast is Faster Risk compared with the state-of-the-art? ( 4.2) |
| Researcher Affiliation | Academia | 1 Duke Univeristy 2 University of British Columbia |
| Pseudocode | Yes | Algorithm 1 Faster Risk(D,k,C,B, ,T,Nm) ! {(w+t, w+t0 , mt)}t |
| Open Source Code | Yes | Code Availability: Implementations of Faster Risk discussed in this paper are available at https://github.com/ jiachangliu/Faster Risk. |
| Open Datasets | Yes | The datasets used are mammo [15], adult, bank, mushroom, COMPAS [21], FICO [17], and Netherlands. All datasets are publicly available from the UCI Machine Learning Repository [20], except COMPAS which is from ProPublica [21], and FICO which is from the FICO Explainable Machine Learning Challenge [17]. |
| Dataset Splits | Yes | For each dataset, we perform 5-fold cross validation and report training and test AUC. |
| Hardware Specification | Yes | Our experiments are run on a Linux machine with 16 Intel Xeon E5-2630 v4 CPUs and 2 NVIDIA Titan X Pascal GPUs. |
| Software Dependencies | Yes | All experiments are conducted on Python 3.8.10. We use numpy 1.21.5 and scipy 1.7.3 for numerical operations, scikit-learn 1.0.2 for baselines, and Gurobi 9.5.1 as the MIP solver. |
| Experiment Setup | Yes | Input: dataset D (consisting of feature matrix X 2 Rn p and labels y 2 Rn), sparsity constraint k, coefficient constraint C = 5, beam search size B = 10, tolerance level = 0.3, number of attempts T = 50, number of multipliers to try Nm = 20. |