- SHAP : A Unified Approach to Interpreting Model Predictions
- HEDGE : Generating Hierarchical Explanations on Text Classification via Feature Interaction Detection
- Evaluation of time series by perturbating subsequences of the data
- Basic idea is to observe the drop in classification quality if perturbating the most important features both i) in isolation and ii) together with the following timestamps
- Temporal Saliency Rescaling
- The main idea behind Temporal Saliency Rescaling is to divide feature importance attribution into two steps: first identify the important time steps and then within the time step the important feature. This should avoid the alleged issue that saliency methods tend to select a full time step as important although only a few features might be relevant
- Unclear how the evaluation in fact works. Seems like the main idea is to just look at precision/recall for perturbed time series
- Benchmarking Deep Learning Interpretability in Time Series Predictions
- The main idea is a network architecture that allows to calculate a saliency map separately for the time and feature domain
- Grad-CAM: Why did you say that? Visual Explanations from Deep Networks via Gradient-based Localization
- Model-Agnostic Explanations
- Deep Inside Convolutional Networks: Visualising Image Classification Models and Saliency Maps
- Visualizing and Understanding Convolutional Networks
- Striving for Simplicity: The All Convolutional Net
- Learning Important Features Through Propagating Activation Differences
- On Pixel-Wise Explanations for Non-Linear Classifier Decisions by Layer-Wise Relevance Propagation
- SmoothGrad: removing noise by adding noise
- Expected gradients
- Better Attributions through regions
- Visualizing the Impact of Feature Attribution Baselines