1 Introduction
2 Related work
Reference | XAI | Domain | Human-centricity | Semantics |
---|---|---|---|---|
Guidotti et al. (2018) | Yes | Explainability of black box models | Yes | No |
Adadi and Berrada (2018) | Yes | XAI for black-box models and trends in XAI research | No | No |
Dosilovic et al. (2018) | Yes | Recent developments in XAI and supervised learning techniques | No | No |
Ehsan and Riedl (2020) | Yes | XAI and the perception of non-expert users | Yes | No |
Islam et al. (2021) | Yes | XAI approaches, Credit default prediction | Yes | No |
Alicioglu and Sun (2022) | Yes | Visual analytics for XAI methods | Yes | No |
Liao and Varshney (2022) | Yes | Human-centered XAI and user experience | Yes | No |
Rong et al. (2022) | Yes | User studies in XAI | Yes | No |
Damfeh et al. (2022) | Yes | Theoretical principles and paradigms for human-centered AI | Yes | No |
Chaddad et al. (2023) | Yes | XAI approaches, Healthcare | Yes | No |
Saeed and Omlin (2023) | Yes | XAI’s challenges and research directions | Yes | No |
Our Survey | Yes | XAI approaches, Event detection | Yes | Yes |
3 Methodology
3.1 Research questions
3.1.1 Human-centric explanations
3.1.2 Explainable event detection
3.1.3 Semantics-based explainable event detection
4 Results
4.1 Human-centric explanations
4.1.1 Explainable AI
4.1.2 Explainable AI techniques
Type of XAI Approaches for Explanations | Description |
---|---|
Ante-hoc/Intrinsic vs. Post-hoc | Intrinsic explainability incorporates explainability directly into their structures. Explainability is achieved by finding large coefficient features that play a significant role in the prediction. Ante-hoc models are more interpretable. Decision trees, K-nearest neighbours, and linear and logistic regression are interpretable. Useful for training a new model for which comprehension is essential. With post-hoc explainability, a second model is required to provide an explanation for the existing one. Support vector machines, ensemble algorithms, and neural networks are examples of models that are intrinsically uninterpretable. The post-hoc explanation can supply an explanation for the intrinsically interpretable models. Useful to leverage already trained or proven machine learning techniques. |
Model-specific vs. Model-agnostic | Post-hoc explanations can be divided into model-specific and model-agnostic explanations. Model-specific is sometimes called white-box explanations because they provide explanations based on the model’s internals. Saliency maps are an example of a model-specific explanation. Saliency maps highlight features that are perceived to influence classification. However, model-specific explanations are designed for certain types of models. The model-agnostic explanation provides explanations that are decoupled from the model. Methods for model-agnostics are partial dependency plots (PDP) and individual conditional expectations. Both methods provide an explanation for the whole model using visual interactions of the model under investigation. |
Surrogate | Another kind of post-hoc explanation that can be used to explain more complicated methods is surrogate. An illustration of a surrogate method that explains ensemble models is the Combined Multiple Models (CMM). |
Global vs. Local | Global explanations try to explain the whole model. Familiar methods used are tree-based models. The local explanation provides explanations based on a region around a single prediction. This is much simpler than global explanations. Examples of popular methods in this category are Local Interpretable Model Explanation (LIME) and SHapley Additive exPlanations. However, SHAP is computationally costlier than LIME. DeepLIFT, GRAD-CAM are used specifically for Deep Learning models explanation. |
4.1.3 Comparison of XAI techniques
XAI Techniques (Ante-hoc/ Post-hoc) | Model Specificity/ Explanation Scope | Idea | Strength | Weakness | Application/ (Target Audience) | Human-centricity | ||||
---|---|---|---|---|---|---|---|---|---|---|
Trustworthiness | Fairness | Transferability | Rationale | Causality | ||||||
LIME (Post-hoc) | Model-agnostic/ Local | Uses surrogate-based explanation to explain a complex model prediction | No need for knowledge of the model’s internal. Offers an interpretable representation. Provides local fidelity | The quality of the explanation is directly proportional to the quality of the surrogate fit. High computational cost | Text and image analysis (Domain Experts) | No | No | No | Yes | No |
SHAP (Post-hoc) | Model-agnostic/ Local or Global | Analyses a feature’s contribution to the model prediction to determine its significance, with features that do not contribute to the prediction receiving zero. | The model’s predictions are broken down additively into components related to specific features. | High computational complexity. SHAP values are symmetrical | Tabular data (AI Layman) | Yes | No | No | Yes | No |
Anchor (Ante-hoc) | Model-agnostic / Local | Finds a decision rule that sufficiently “anchors” a prediction | Can have high coverage and high precision. Can be applied to different domains | Computationally intensive. Unbalanced classification problem leads to trivial decision rules | Text, image, and tabular data analysis (Domain Experts) | No | No | No | Yes | Yes |
GraphLIME (Post-hoc) | Non-linear model-agnostic/ Local | Computes K most representative features to explain its prediction and produces a non-linear interpretable model from the N-hop neighbourhood of the node. | Can filter useless features and select informative features as explanations | Provides explanations for only node features as it ignores graph structures, such as nodes and edges. Not suitable for graph classification problems | Graph analysis (Domain Experts) | No | No | No | Yes | No |
LRP (Post-hoc) | Model specific/ Local or Global | Takes advantage of the network structure and redistributes the explanations from the model’s output to the input layer by layer. | Makes use of additional features of the model internal to provide a better explanation. High computational efficiency | Adaptation to novel model architectures is difficult | NLP, computer vision, meteorology, games, video, morphing, EEG/fMRI analysis (Domain Experts) | No | No | No | Yes | No |
DTD (Post-hoc) | Model-agnostic/ Local | Employs the first order of Taylor expansion to redistribute the neural network’s output to the input variables layer-wise. | Computationally efficient | Adaptation to novel model architectures is difficult | Image analysis (Domain Experts) | No | No | No | Yes | No |
PDA (Post-hoc) | Model-agnostic/ Local | Measures the change in prediction when the feature is unknown to determine its relevance. | Can map uncertainty in model prediction to model inputs | Computationally expensive. Can suffer from saturated classifiers | Image analysis (Domain Experts) | No | No | No | Yes | No |
TCAV (Post-hoc) | Model-agnostic/ Global or Local | Explains how neural activations affect the absence or presence of a user-specific concept. | Usable by users without prior experience with machine learning | Not suitable for tabular, text data or shallower neural networks | Concept sensitivity in image, fairness analysis (Domain Experts) | No | No | No | Yes | No |
XGNN (Post-hoc) | Model-agnostic/ Local | Applies reinforcement learning to obtain important graph generation for GNN model prediction | Operates on the model level, and no need to provide individual-level explanations | Absence of ground truth results in non-concrete explanations | Graph classification (AI Layman) | Yes | No | No | Yes | Yes |
ASV (Post-hoc) | Model-agnostic/Global | Uses cause-effect relationship to redistribute attribution of features in a manner that the source feature has higher attribution that provides an effect on the model’s predictions as well as the other dependent features | Only the features that are consistent with the causal features are considered. Does not require model retraining for feature selection | Requires domain knowledge | Fairness analysis (Domain Experts) | No | Yes | No | Yes | Yes |
Break-Down (Post-hoc) | Model-agnostic/Local | Uses greedy heuristics to identify and visualise the model’s interactions to determine the final attributions based on single ordering | Variable contributions are provided in a concise way | The part of the prediction attributed to a variable depends on the order in which one sets the values of the explanatory variables | Tabular dataset (AI Layman) | No | No | No | Yes | Yes |
Shapley Flow (Post-hoc) | Model-agnostic/Local | Uses a dependency structure between features for an explanation, as in ASV. However, attributions are assigned to the relationship between features, unlike ASV, where attributions are assigned to features themselves | Have a lot of information about the relationship’s structure between features and explanation boundaries | Requires familiarity with the structure of dependencies and the knowledge of the background case, that is, reference observation | Graph | No | No | No | Yes | Yes |
Textual Explanations of Visual Models (Post-hoc) | Model-specific/ | Finds the discriminative characteristics to generate explanations | More straightforward to analyse and verify than attribution maps | There is no way to verify that the generated explanation matches the domain expertise. Artefacts of data can negatively influence performance and quality of explanation | Text and image | No | No | No | Yes | No |
Integrated Gradients (post-hoc) | Model-agnostic/ Local | Uses sensitivity and implementation invariance properties to achieve the model’s prediction explanation | Computationally efficient. Makes use of gradient information at a few specific locations. | Requires baseline observation. Suitable for only differential models. Suffers from gradient shattering problem | Text and image analysis | No | No | No | Yes | No |
Causal models | Model-agnostic/ Global | Uses reinforcement learning theory for the counterfactual explanation that provides causal chains up until the rewards-receiving state | The “what”, “how”, and “why” questions are taken care of | Applicable to a finite domain | Text | No | No | No | Yes | Yes |
Meaningful Perturbations | Model-agnostic/ Local | Generates explanations based on the model’s reaction to a perturbed input sample | Very flexible | Computationally expensive | Text, image, and tabular data analysis | No | No | No | Yes | No |
EXplainable Neural-Symbolic Learning | Model-agnostic/ Local | Aligns the symbolic knowledge of domain experts (the Knowledge Graph) with the neural network’s explanations, which correspond to the human classification method. | Boosts explainability and sometimes performance | Requires domain-specific knowledge | Text and image analysis | No | Yes | No | Yes | Yes |
Saliency Maps (Post-hoc) | Model-specific/ Local | calculates feature importance on gradients, visualises and emphasises important pixels that influence the final CNN decision | Can analyse the image regions that stood out across the whole dataset | Saturation problem | Image analysis (End users) | Yes | No | No | Yes | No |
CAM (Post-hoc) | Model-agnostic/ Local | A gradient-based explanation approach that makes use of global average pooling for class activation maps in CNN | Identification of important regions in an image. Can explain graph classification models. | It requires a specific CNN architecture without any fully connected layers. Cannot be applied directly to the classification of nodes. | Image analysis | No | No | No | Yes | No |
DeepLift (Post-hoc) | Model-agnostic/ Local | Propagates a reference (neutral or default) input to obtain a reference output. Allocates importance scores using the difference between the actual output and the reference | Can reveal dependencies | It is not implementation invariant, i.e. two identical models with different internal wiring could produce different outputs | Image analysis | No | No | No | Yes | Yes |
Bayesian Rule Lists (Ante-hoc) | Model specific/ Global | Creates IF-THEN rule sets to provide an explanation | Reduces model space by using pre-mined rules | Rules can overlap | Text (End users) | Yes | No | No | Yes | Yes |