Sparse, Efficient and Explainable Data Attribution with DualXDA
Authors: Galip Ümit Yolcu, Moritz Weckbecker, Thomas Wiegand, Wojciech Samek, Sebastian Lapuschkin
TMLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In extensive quantitative analyses, we demonstrate that Dual DA achieves high attribution quality, excels at solving a series of evaluated downstream tasks, while at the same time improving explanation time by a factor of up to 4,100,000 compared to the original Influence Functions method, and up to 11,000 compared to the method s most efficient approximation from literature to date. We evaluate Dual DA in terms of quality and sparsity of explanations, and runtime as well as memory efficiency. In total, we compare nine different methods on seven quantitative metrics across three different datasets and models, in the most extensive quantitative evaluation of prominent DA methods so far. |
| Researcher Affiliation | Academia | Galip Ümit Yolcu Fraunhofer Heinrich Hertz Institute Moritz Weckbecker Fraunhofer Heinrich Hertz Institute Thomas Wiegand Fraunhofer Heinrich Hertz Institute Technische Universität Berlin BIFOLD Berlin Institute for the Foundations of Learning and Data Wojciech Samek EMAIL Fraunhofer Heinrich Hertz Institute Technische Universität Berlin BIFOLD Berlin Institute for the Foundations of Learning and Data Sebastian Lapuschkin EMAIL Fraunhofer Heinrich Hertz Institute Technological University Dublin |
| Pseudocode | No | The paper describes methods step-by-step and provides mathematical derivations, but does not include any explicitly labeled pseudocode or algorithm blocks. |
| Open Source Code | Yes | The implementation of our methods, as well as the full experimental protocol, is available on github1. 1 https://github.com/gumityolcu/Dual XDA |
| Open Datasets | Yes | MNIST dataset (Le Cun et al., 2010), CIFAR-10 (Krizhevsky & Hinton, 2009) and Res Net-50 (He et al., 2016) trained on the Animals with Attributes 2 (Aw A2) dataset (Xian et al., 2018), Image Net dataset (Deng et al., 2009), AG News dataset for news topic classification (Zhang et al. (2015), available on Huggingface6). |
| Dataset Splits | Yes | For each dataset and model, we calculate the attribution score of all training datapoints for the predicted class for 2,000 randomly selected test points... We report the test loss when retrained on 10% of the data as well as a weighted average p) ℓCS test(p) P where ℓCS test(p) is the test loss when retrained only on the most relevant p% of the training data and we use steps of 10%, i.e. P = {10, . . . , 90}. We report the test loss when retrained without the 10% most influential data as well as a weighted average p) ℓDP test(p) P where ℓDP test(p) is the test loss when retrained without the most relevant p% of the training data and we again choose P = {10, . . . , 90}. |
| Hardware Specification | Yes | All experiments have been executed on a single NVIDIA A100 Tensor Core-GPU. All experiments have been run on a NVIDIA H200 Tensor Core GPU. |
| Software Dependencies | No | To implement Dual DA in Python, we have modified4 the multiclass SVM implementation from the Python package scikit-learn (Pedregosa et al., 2011). Py Torch (Paszke et al., 2019) framework. For Arnoldi, we relied on the captum (Kokhlikyan et al., 2020) implementation. Within the LRP framework, different formulas, or rules , exist to distribute the relevance from the output back to the input space. For our experiments, we use the recommended rule composite Epsilon Plus Flat, as implemented in the zennit toolkit (Anders et al., 2021). |
| Experiment Setup | Yes | For MNIST, we use a 6-layer convolutional neural network with 0.001 a learning rate of 0.001. For CIFAR-10, we train a Res Net-18 (He et al., 2016) with random cropping and flipping as data augmentations, with a learning rate of 0.0003. We further include a weight decay term with a coefficient of 0.01 as part of the loss term. For Aw A2, we have use a learning rate of 0.001 and augment the data with random horizontal flipping of the images as data augmentation for training a Res Net-50. For Arnoldi explanations, we have used 128 as the number of random projections and 150 as the Arnoldi space dimensionality. We have used 10,000 randomly sampled datapoints from the training dataset to estimate the Hessian matrix. For TRAK... we use 2,048 random projections. The training durations for the SVM surrogate models are as follows: 18 minutes 27 seconds for C = 10 1, 48 seconds for C = 10 3, and 32 seconds for C = 10 5. |