Philine Bommer,Marlene Kretschmer,Anna Hedström,Dilyara Bareeva,Marina M.-C. Höhne
Abstract:Explainable artificial intelligence (XAI) methods shed light on the predictions of machine learning algorithms. Several different approaches exist and have already been applied in climate science. However, usually missing ground truth explanations complicate their evaluation and comparison, subsequently impeding the choice of the XAI method. Therefore, in this work, we introduce XAI evaluation in the climate context and discuss different desired explanation properties, namely robustness, faithfulness, randomization, complexity, and localization. To this end, we chose previous work as a case study where the decade of annual-mean temperature maps is predicted. After training both a multi-layer perceptron (MLP) and a convolutional neural network (CNN), multiple XAI methods are applied and their skill scores in reference to a random uniform explanation are calculated for each property. Independent of the network, we find that XAI methods Integrated Gradients, layer-wise relevance propagation, and input times gradients exhibit considerable robustness, faithfulness, and complexity while sacrificing randomization performance. Sensitivity methods -- gradient, SmoothGrad, NoiseGrad, and FusionGrad, match the robustness skill but sacrifice faithfulness and complexity for randomization skill. We find architecture-dependent performance differences regarding robustness, complexity and localization skills of different XAI methods, highlighting the necessity for research task-specific evaluation. Overall, our work offers an overview of different evaluation properties in the climate science context and shows how to compare and benchmark different explanation methods, assessing their suitability based on strengths and weaknesses, for the specific research problem at hand. By that, we aim to support climate researchers in the selection of a suitable XAI method.
What problem does this paper attempt to address?
### What problem does this paper attempt to solve?
This paper aims to solve the problem of selection and evaluation of Explainable Artificial Intelligence (XAI) methods in climate science. Specifically, it focuses on the following aspects:
1. **Evaluation and comparison of XAI methods**:
- In the field of climate science, existing XAI methods lack a clear "ground truth explanations", which makes the evaluation and comparison of different XAI methods complicated.
- The paper introduces an evaluation framework for evaluating the performance of different XAI methods in the context of climate science.
2. **Key properties of XAI methods**:
- The paper discusses the key properties that XAI methods should possess, including **robustness**, **fidelity**, **randomization**, **complexity** and **localization ability**. These properties help to measure the reliability and effectiveness of XAI methods when explaining the predictions of machine - learning models.
3. **Case study**:
- Through a specific case study, the paper shows how to apply multiple XAI methods to explain the prediction results of multi - layer perceptron (MLP) and convolutional neural network (CNN) for annual average temperature maps.
- The study finds that different XAI methods have differences in performance in terms of robustness, fidelity, complexity and randomization, and these differences are related to the network architecture.
4. **Providing guidance**:
- The paper provides guidelines for climate researchers to select appropriate XAI methods, helping them to make selections according to the advantages and disadvantages of XAI methods in specific research tasks.
### Formula summary
- **Robustness evaluation formula**:
\[
q_{LLE,m}=\max_{x + \delta\in N_{\epsilon}(x)}\frac{\|\Phi_m(f,c,x)-\Phi_m(f,c,x + \delta)\|_2}{\|x-(x + \delta)\|_2}
\]
\[
q_{AS,m}=E_{x+\delta\in N_{\epsilon}(x)}\left[\frac{\|\Phi_m(f,c,x)-\Phi_m(f,c,x + \delta)\|}{\|x\|}\right]
\]
- **Fidelity evaluation formula**:
\[
\hat{q}_{ROAD,m,i}=\frac{1}{N}\sum_{n = 1}^N1_{c_n}(c_{pred,n})
\]
where,
\[
1_{c_n}(c_{pred,n})=\begin{cases}
1&\text{if }c_n = c_{pred,n}\\
0&\text{otherwise}
\end{cases}
\]
Through these formulas, the paper can quantify the performance of different XAI methods in various key properties, thus providing a scientific basis for climate science researchers to help them select the most appropriate XAI method.