Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
SHAP values via sparse Fourier representation
Authors: Ali Gorji, Andisheh Amrollahi, Andreas Krause
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | 5 Experiments We assess the performance of our algorithm, FOURIERSHAP, on four different real-world datasets of varying nature and dimensionality. ... In Figure 1 we plot the accuracy of the Fourier function approximation as measured by the R2-score for different values of k... In Figure 2 we plot this trade-off. |
| Researcher Affiliation | Academia | Ali Gorji ETH Zürich, Switzerland EMAIL Andisheh Amrollahi ETH Zürich, Switzerland EMAIL Andreas Krause ETH Zürich, Switzerland EMAIL |
| Pseudocode | No | The paper describes the proposed method and its steps using prose and mathematical equations (e.g., Equation 5) but does not include a distinct block labeled 'Pseudocode' or 'Algorithm'. |
| Open Source Code | Yes | The full code base is provided as supplementary material to the paper. We refer the reader there for more details. Here we give a high level overview of the four versions we implemented and experimented with: ... The code for running the experiments and the implementations of all modules are open-sourced and is publicly available at https://github.com/andisheh94/fouriershap. |
| Open Datasets | Yes | We assess the performance of our algorithm, FOURIERSHAP, on four different real-world datasets of varying nature and dimensionality. Three of our datasets are related to protein fitness landscapes [Poelwijk et al., 2019, Wu et al., 2016, Sarkisyan et al., 2016] and are referred to as Entacmaea (dimension n = 13), GB1 (n = 80), and av GFP (n = 236) respectively. The fourth dataset is a GPU-tuning [Nugteren and Codreanu, 2015] dataset referred to as SGEMM (n = 40). |
| Dataset Splits | Yes | We fit random forests of maximum depths ranging from 3 to 8 with 20 estimators on 90% of all four datasets used in the black-box setting. |
| Hardware Specification | Yes | We run all experiments on a machine with one NVIDIA GeForce RTX 4090 GPU, on servers with Intel(R) Xeon(R) CPU E3-1284L v4 @ 2.90GHz, restricting the memory/RAM to 20 GB, which was managed with Slurm. |
| Software Dependencies | No | The paper mentions using JAX [Bradbury et al., 2018], sklearn [Pedregosa et al., 2011], and cat-boost [Dorogush et al., 2018] but does not provide specific version numbers for these software dependencies, which are required for full reproducibility according to the specified criteria. |
| Experiment Setup | Yes | For the Entacmaea and SGEMM datasets, we train fully connected neural networks with 3 hidden layers containing 300 neurons each. The network is trained using the means-squared loss and ADAM optimizer with a learning rate of 0.01. For GB1 we train ensembles of trees models of varying depths using the random forest algorithm and for av GFP we train again, ensembles of trees models of varying depths using the cat-boost algorithm/library [Dorogush et al., 2018]. All other setting are set to the default in both cases. |