Prediction via Shapley Value Regression
Authors: Amr Alkhatib, Roman Bresson, Henrik Boström, Michalis Vazirgiannis
ICML 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate both the predictive performance of Via SHAP and the feature importance attribution with respect to the true Shapley values. This section begins with outlining the experimental setup. Then, the predictive performance of Via SHAP is evaluated. Afterwards, we benchmark the similarity between the feature importance obtained by Via SHAP and the ground truth Shapley values. We also evaluate the predictive performance and the accuracy of Shapley values on image data. Finally, we summarize the findings of the ablation study. |
| Researcher Affiliation | Academia | 1 Orebro University, School of Science and Technology, Sweden 2KTH Royal Institute of Technology, School of Electrical Engineering and Computer Science, Sweden 3 Ecole Polytechnique, IP Paris, France. Correspondence to: Amr Alkhatib <EMAIL>. |
| Pseudocode | Yes | Algorithm 1 VIASHAP Data: training data X, labels Y , scalar β Result: model parameters θ Initialize V : Via SHAP(ϕVia(x; θ)) while not converged do L 0 for each x X and y Y do sample S p(S) ˆy V(x) Lpred prediction loss(ˆy, y) Lϕ Vy(x S) Vy(0) 1 S ϕVia y (x; θ) 2 L + Lpred + β Lϕ end Compute gradients θL Update θ θ θL end |
| Open Source Code | Yes | The source code is available here: https://github. com/amrmalkhatib/Via SHAP |
| Open Datasets | Yes | We employ 25 publicly available datasets in the experiments, each divided into training, validation, and test subsets 1. ... 1The details of the datasets are available in Table 19 ... For image experiments, we use the CIFAR-10 dataset (Krizhevsky et al., 2014). |
| Dataset Splits | Yes | We employ 25 publicly available datasets in the experiments, each divided into training, validation, and test subsets 1. The training set is used to train the model, the validation set is used to detect overfitting and determine early stopping, and the test set is used to evaluate the model s performance. ... Table 19: The dataset information. # Features # Classes Dataset Size Train. Set Val. Set Test Set Open ML ID |
| Hardware Specification | Yes | The experiments were conducted using an NVIDIA Tesla V100f GPU and 16 cores of an Intel Xeon Gold 6338 processor. |
| Software Dependencies | No | The paper mentions various frameworks and models used (e.g., KANs, MLPs, ResNet50, U-Net) and references implementations like efficient-kan and fast-kan, but does not provide specific version numbers for underlying software dependencies like Python, PyTorch, or CUDA. |
| Experiment Setup | Yes | The four implementations were trained with the β of (7) set to 10 and used 32 sampled coalitions per instance. The above hyperparameters were determined in a quasi-random manner. ... During data preprocessing, categorical feature categories are tokenized with numbers starting from one, reserving zero for missing values. We use standard normalization so the feature values are centered around 0. Via SHAP can be trained using the baseline removal approach or marginal expectations as a value function. |