Explainable AI (XAI): Core Ideas, Techniques, and Solutions

Probably not a substantial claim: may not pass S3 (load-bearing).

In the cancer dataset example, the Worst Concave points feature initially has an increasing negative influence and then remains neutral.

“In our example,”

Probably not a substantial claim: may not pass S1 (asserted), S3 (load-bearing), S4 (based). May combine two claims.

In the cancer dataset ICE plot, the worst concave points plot is decreasing for lower values below 0.10, which is a factor for higher target variable malignant cases.

“The”

Probably not a substantial claim: may not pass S3 (load-bearing).

During training of sequential models, only one GPU is utilized at a time because it waits for results from previous or subsequent GPU.

“For example during training of the sequential models, only one GPU is utilized at a time as it has to wait for results from the previous/ subsequent GPU.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Between 0.10 and 0.20 the ICE graph decreases, and after 0.21 it is constant with lower y value benign cases.

“Between 0.10 and 0.20 the graph decreases and after 0.21 it is constant (has lower y value -benign cases).”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

In the cancer dataset SHAP example, the predicted value was -3.89 and the base value was 1.837.

“We predicted -3.89, whereas the basevalue is 1.837.”

Probably not a substantial claim: may not pass S3 (load-bearing).

In the shirt example, positive and negative attributions are equal in number.

“The second example (bottom part of Figure 20) is an image of a shirt where there is an equal number of positive and negative attributions.”

May combine two claims.

In the Iris ALE example, greater petal length increases chance of Virginica species, and lesser petal length increases chance of Setosa species.

“It can also be observed that the more the petal length is, the more is the chance of the flower being from the Virginica species and the less the petal length is, the more is the chance of it being from the Setosa species.”

Probably not a substantial claim: may not pass S3 (load-bearing).

In the shirt example, pixels around the collar and sleeves are the largest positive attributions.

“The pixels around the collar and the sleeves are the biggest positive attributions.”

Probably not a substantial claim: may not pass S3 (load-bearing).

In the heart disease CEM example, the original prediction was 0 and was altered to 1 after applying CEM with pertinent negative.

“The original prediction was 0 which is altered to 1 after applying CEM with pertinent negative.”

Probably not a substantial claim: may not pass S3 (load-bearing).

In the cancer dataset SHAP example, mean concave points had a high effect 0.04079 increasing the prediction.

“Though the mean concave points value has high effect (0.04079) increasing the prediction.”

Probably not a substantial claim: may not pass S3 (load-bearing).

In the Iris dataset ALE example, the three lines for Petal Length overlap at the 3.8 cm mark.

“Looking at the plot for Petal Length (bottom right plot), it can be observed that the three lines overlap at the 3.8 cm mark.”

Probably not a substantial claim: may not pass S3 (load-bearing).

In the shoe example, negative attributions are negligible.

“On the other hand, the negative attributions are negligible for this particular instance.”

Probably not a substantial claim: may not pass S3 (load-bearing).

In the shoe example, the lining, collar, and back of the shoe are the main pixels influencing the model's decision.

“The lining, collar and back part of the shoe are the main pixels that influence the decision of the model.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

In the heart disease global SHAP example, ca is the feature that influences prediction the most, followed by thal.

“Thus, ca is the feature that influences prediction the most; followed by thal and so on.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

In the heart disease local SHAP example, the highest influence for increasing the prediction value is ca, and for decreasing it is sex.

“Here, the highest influence is of ca for increasing the prediction value and of sex for decreasing the value.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

In the Fashion MNIST shoe example, the attribution map contains both positive and negative attributions, with green marking positive and purple marking negative.

“The first example (top part of Figure 20) is is an image of a shoe.The attributions section shows a melange of positive and negative attributions together.It can be observed from the bar on the right side that green signifies positive attributions and purple signifies negative attributions.”

Probably not a substantial claim: may not pass S3 (load-bearing).

In the heart disease counterfactual example, reducing cholesterol leads to a decrease in heart disease intensity.

“the second counterfactual on the list indicates that reduction of cholesterol leads to a decrease in the intensity of heart disease.”

Probably not a substantial claim: may not pass S3 (load-bearing).

In the cancer dataset SHAP example, the biggest impact came from worst area 989.5.

“The biggest impact comes from worst area i.e. 989.5.”

Probably not a substantial claim: may not pass S3 (load-bearing).

In the heart disease counterfactual example, sex, age, and chest-pain type cannot be changed and must be fixed in each counterfactual.

“The sex, age or the type of chest-pain (field cp) of a person suffering from a heart disease cannot be changed.Therefore, these features have to be fixed in each of the counterfactuals.”

May combine two claims.

XAI models can present the reasoning behind a result, which helps identify the source of discrimination and address it.

“Unlike black box models, XAI models can present the reasoning behind the result, which can help identify the source for the discrimination and address it accordingly.”

Probably not a substantial claim: may not pass S4 (based).

An ML pipeline that provides accurate predictions lacks the phases of understanding and explaining.

“While an ML pipeline can provide accurate predictions, it lacks two important phases: understanding and explaining.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Feature importance shows the comparative significance of input features for a specific model prediction.

“For example, feature importance (discussed in Section 6.1), a common type of explanation, shows the comparative significance of input features for any specific model prediction(s).”

Probably not a substantial claim: may not pass S4 (based). May combine two claims.

The architecture of black-box models is hard to decipher, and it is unclear how important any feature is or how it interacts with other features.

“The architecture of these models is hard to decipher, as it is not clear how important a role any given feature plays in the prediction model or how it interacts with other features.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Global interpretation methods analyze a model's overall behavior and assign importance to variables, their dependencies, and interactions.

“Global interpretation methods involve an overall analysis of a model and its general behavior.The process of defining variables, their dependency, and interactions goes alongside with the process of assigning importance to these components.”

• cites [44]Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

Linear models are white-box models because their mechanism is transparent and simple.

“A linear model is a white-box model because its mechanism is transparent and simple (as opposed to a black-box whose mechanism is not readily understood).”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

In a fully connected neural network, tracing output features to a specific causative input feature remains a challenge.

“For example, in a fully connected neural network, tracing the output features rendered by a model against a specific causative input feature remains a challenge.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Higher accuracy requires more complex and expressive models.

“Therefore, for higher accuracy, we require more complex and expressive models.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Black-box models include neural networks and complex ensembles such as gradient boosting models based on decision trees.

“Black box models such as neural networks or complex ensembles of much lower complexity (like a gradient boosting model based on decision-trees).”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Datasets often contain inherent discrimination.

“Datasets often contain inherent discrimination.”

May combine two claims.

AutoML turns the entire data pipeline into a black box, preventing users from engineering or selecting features or seeing the model's decision-making process.

“AutoML has made explainability more important, as the entire data pipeline is convered into a black box (as opposed to just the machine learning model) [27].When using AutoML, a user does not have the ability to engineer or select features, or has visibility into a model's decision-making process.”

• cites [27]Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing).

ICE curves are far more intuitive to understand than PDP.

“They are far more intuitive to understand when compared to PDP [44].”

• cites [44]Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing).

Symbols are closer to how humans understand than vectors or numerical encodings used in neural networks.

“Symbols are closer to how humans understand, compared to vectors or numerical encodings as seen in neural networks.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Logistic regression is one of the most interpretable linear machine learning models for certain classes of events.

“Similarly, Logistic Regression is one of the most interpretable linear ML models for certain class of events.”

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing).

Deep understanding of cognitive science will benefit the field of neuro-symbolic AI, particularly for integrating neuro-symbolic approaches and designing explainable systems closer to human understanding.

“In fact, deep understanding of cognitive science will most certainly benefit the field of NS-AI particularly for the integration of neuro-symbolic approaches [6] to develop a mature and robust integrated model and consequently, for the design of explainable systems that are closer to how humans understand, think, and communicate.”

• cites [6]Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

Generalized additive models offer a trade-off between simple interpretable models like logistic regression and more complex models like neural networks.

“GAMs offer a trade-off between simple, interpretable models such as logistic regression and more complex, sophisticated models such as neural networks, which (usually) offer better accuracy and predictive power as compared to simple models.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Local interpretation methods analyze individual predictions and decisions to clarify why the model suggested a particular course of action.

“Involves an analysis of individual predictions and decisions made by the model, to clarify why the model suggested a particular course of action.”

• cites [44]Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Tree interpreter decomposes every prediction result into a sum of feature contributions and bias, enabling better interpretation of how a feature led to a prediction.

“It decomposes every prediction result to a sum of feature contributions and bias.This enables a better interpretation of how a feature led to a particular prediction.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Too many ICE curves can overcrowd a plot and prevent assessment.

“First, too many ICE curves could lead to an overcrowded plot without the ability to assess anything.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Neural networks usually offer better accuracy and predictive power than simple models.

“which (usually) offer better accuracy and predictive power as compared to simple models.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Linear model coefficients can be unstable when features are multicollinear.

“Furthermore, the linear model coefficients could be unstable in case of multi-collinearlity (multiple correlated features).”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Post-hoc explanation schemes are used for single-layer and multilayer neural networks, convolutional neural networks, and recurrent neural networks.

“These post-hoc explanation schemes are being used for both single/ multilayer Neural Networks, Convolutional Neural Networks and Recurrent Neural Networks.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

The InTrees package converts a complex tree ensemble into a rule-based learner called STEL.

“The 'InTrees' (Interpretable Trees) package coverts a complex tree ensemble into a rule based learner known as STEL.”

Probably not a substantial claim: may not pass S2 (checkable).

Reasoning is critically needed to understand the why and how behind a particular outcome.

“However, to achieve this explainability, the authors argue the critical need for reasoning to understand the why and how behind a particular outcome [14].”

• cites [14]Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

Synchronous training usually takes more time because it waits for idle nodes.

“Synchronous training usually takes more time as it requires waiting for idle nodes.”

May combine two claims.

Neural systems excel at learning, while symbolic approaches can explain what has been learned.

“Thus, where neural systems excel at learning, symbolic approaches can be employed to explain what has been learnt.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Ring-allreduce usually achieves better scalability due to efficient network bandwidth use.

“The ring-allreduce [10] usually achieves better scaleability due to efficient use of network bandwidth.”

• cites [10]Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

Histograms are much less flexible than kernel density estimates for data visualization.

“Although Data Visualization Plots give a good data explainability, plots such as Histogram are much less flexible than KDEs.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Permutation importance is considered computationally expensive.

“It is considered as a computationally expensive technique.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Dimensionality reduction techniques such as PCA, ICA, Isomap, t-SNE, LDA, UMAP, LDS, LLE, and autoencoders can improve visualization and interpretation.

“Dimensionality reduction techniques such as PCA, ICA, Isometric Mapping (Isomap), t-sne, LDA, UMAP, LDS, LLE, and Autoencoders can be used to improve visualisation and interpretation.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Training large models and data volumes requires distributing ML workload across multiple machines and aggregating results because of single-machine limits in computing power and storage.

“To train for large models and data volumes, the limitation of the computing power and storage of a single machine calls for distributing the machine learning workload across multiple machines and aggregating the results for a coherent model.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Tree ensembles provide explainability similar to decision trees while offering stronger learning during training.

“Tee ensembles can be understood as a bunch of decision trees whose results are combined thus providing the support for explainability same as decision trees but with the benefit of stringer learning during training.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Small ML models lack representation capability for complex data patterns.

“Small ML models lack the representation capability for expressing complex data patterns.”

May combine two claims.

Symbolic systems lack scalability and cannot handle noisy data, which has limited their use.

“However, the lack of scaleability of symbolic systems and their inability to handle noisy data has limited their use.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

GAMs are interpretable but do not fit every type of data.

“Although GAMs are interpretable, it comes at an cost of not fitting every type of data.”

May combine two claims.

Linear model coefficients do not guarantee explainability in practice when the input is high-dimensional; explainability may require limiting the number of input features through regularization.

“However, some techniques viz.linear model coefficients do no ensure explainability in practice due to the reason that implementing them on a high dimensional input may not be explainable.It could only be possible if the number of input features are limited using regularization.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

A single decision tree may not yield optimal performance, so several decision trees can be combined to get an optimal performance model.

“Since a single decision tree may not be enough to yield optimal performance, several decision trees can be combined together to get an optimal performance model.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Black-box models include tree ensembles, support vector machines, and various neural networks.

“Black Box models include Tree Ensembles, Support Vector Machines (SVM) as well as a variety of neural networks.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Subsymbolic approaches based on artificial neural networks are growing in popularity because they are robust against noisy data and can perform complex tasks not otherwise manually possible.

“Subsymbolic approaches, such as those based on artificial neural networks, are growing in popularity due to their robustness against noisy data and ability to perform complex tasks not otherwise manually possible.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

Asynchronous SGD provides better model generalization capability and is fast in training.

“However, it provides better model generalization capability and is fast in training.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Ontologies can perform deductive reasoning to derive new, deeper knowledge from existing knowledge in a traceable manner.

“Ontologies can thus perform deductive reasoning to derive new, deeper knowledge from existing knowledge in a traceable manner.”

May combine two claims.

In asynchronous SGD, delayed gradients create noise in the global model and delay convergence speed.

“In Asynchronous SGD, the delayed gradients create noise in the global model and delay the convergence speed.”

• cites [11]Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

GAMs are explainable because feature value distributions can be plotted in 2D for domain experts to interpret.

“GAMs are explainable in the sense that the distributions of values for any given feature function can be plotted in 2D, which a domain expert can easily interpret.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Parameter server can suffer from network congestion at the server side due to model aggregation.

“PS can suffer from network congestion at the server side due to the model aggregation approach.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Overfitting is unlikely in generalized additive models because prediction functions are regularized.

“Over fitting is unlikely in GAMs due to the regularization of prediction functions.”

May combine two claims.

Combining multiple decision trees increases model complexity and makes model behavior harder to understand.

“However, the model's complexity increases due to multiple decision trees.As a result, it becomes a lot more difficult to understand model behaviour.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

GAMs are more flexible than regression.

“A generalized additive model (GAM) is much more flexible than Regression.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

Decision tree visualizations are intuitive and reveal all decisions during model training, supporting interpretability and explainability.

“Decision Tree visualizations are intuituve and reveal all decisions during model training, thus supporting interpretability and explainability.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Parameter aggregation between worker nodes impacts loss in model training quality and convergence speed.

“Parameter aggregation between worker nodes impacts several system metrics: loss in model training quality, convergence speed.”

May combine two claims.

Ontologies enable reasoning for both deducing new knowledge and causal inference to understand cause and effect behind outcomes.

“Ontologies enable reasoning not only to deduce new knowledge and facts but also for causal inference to understand the cause and effect behind the outcomes, such as through transitivity [23].”

• cites [23]Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

Parameter server is the most prominent approach for centralized aggregation of data parallel training.

“Parameter sever (PS) [37] is most prominent for centralized aggregation of data parallel training.”

• cites [37]Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

Support vector machines work well in high-dimensional spaces and when the number of dimensions exceeds the number of samples.

“It works well in high dimensional spaces and scenarios where number of dimensions are greater than the number of samples.”

May combine two claims.

Mao et al.'s Neuro-Symbolic Concept Learner can be trained on 10% of the original dataset while being explainable.

“Mao et al. [41] present a Neuro-Symbolic Concept Learner, which incorporates symbolic reasoning for better interpretation of visual concepts, and can be trained on less data (10% of the original dataset) while being explainable.”

• cites [41]Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Kernel density estimates provide estimates of an unknown density function and allow adjustment of bandwidth and kernel shape and size.

“The latter give estimates of an unknown density function wherein we can not only differ the bandwidth, but also implement kernels with various shapes and sizes.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

DiCE identifies the most influential feature for a prediction and recommends feature modifications needed to change the result.

“DiCE not only provides the most influential feature for the prediction but also recommends the feature modifications needed for the result.”

• cites [45]Not checked yet

  Not checked in the free run.

May combine two claims.

ICE unravels heterogeneous effects but makes it difficult to find the average.

“ICE unravels the heterogeneous effects but makes it difficult to find the average.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

Compared to Shapley values, Kernel SHAP provides computational efficiency and accurate approximation in higher dimensions.

“Compared to Shapley values, Kernel SHAP provides computational efficiency and accurate approximation in higher dimensions.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

Counterfactuals guided by prototypes is more advanced and accurate than counterfactual, and works on black-box models.

“Counterfactuals guided by prototypes is more advanced and accurate, and works on black-box models.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

The linear approximation assumption in LIME is correct to some extent when only a small region around the data sample is considered.

“This assumption is correct to some extent when only a small region around the data sample is considered.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

In the shoe example, the shoe has more positive than negative attributes because it is unique among clothing items.

“The shoe is unique compared to other clothing items, and hence, it has a lot more positive attributes than negatives.”

Probably not a substantial claim: may not pass S3 (load-bearing).

DiCE allows users to specify feature weights to indicate relative difficulty of modifying features, with higher weight meaning more difficult to modify.

“DiCE allows input of relative difficulty by specifying 'feature weights'.A higher feature weight means that the feature is difficult to modify than that of others.”

• cites [45]Not checked yet

  Not checked in the free run.

May combine two claims.

SHAP values provide local transparency and interpretability, allowing each observation to have its own SHAP value set.

“An important advantage of SHAP value lies in its transparency and interpretability locally, which implies that each observation can have its own SHAP value set.”

• cites [40]Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Crunching large background data still incurs high computational costs.

“However, crunching large background data still incurs high computational costs.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

ALE Plots are preferred over PD Plots because they produce optimal results despite correlation between features and are less computationally expensive.

“They are preferred over PD Plots, as they produce optimal results in spite of correlation between features and are less computationally expensive.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

Shapley Value provides local and global interpretation including explanation with theoretical inferences.

“Shapley Value provides local and global interpretation including explanation with theoretical inferences.”

Probably not a substantial claim: may not pass S3 (load-bearing).

In the heart disease global SHAP example, high thalach indicates no heart disease.

“However, it is the opposite for some features: High thalach will indicate no heart disease.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

An interpretable surrogate model may be close for one subset but broadly disparate for a different subset of a dataset.

“It is possible that an interpretable model is close for one subset but broadly disparate for a different subset of a dataset.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Anchors support model-agnostic approaches for classification models of text, image or tabular data.

“Anchors support model-agnostic approaches for classification models of text, image or tabular data.”

May combine two claims.

Kernel SHAP method applies a fast kernel approximation and may resolve the computational expense of Shapley Value.

“This problem may be resolved with the recently developed kernel SHAP method which applies a fast kernel approximation.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Alibi contains a wide range of algorithms, most focusing on black-box models and local interpretation.

“The library consists of a wide range of algorithms, most of which focus on black box models and local interpretation i.e. interpretation for the prediction to be explained at a particular instance.”

• cites [34]Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

In the heart disease CEM example, changes in features cp, ca, and thal should be eliminated to retain the original prediction as 0 because they are responsible for flipping the prediction class.

“Hence, changes in these features should necessarily be eliminated to retain the original prediction as 0 as they are responsible for flipping the prediction class.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

PN works similarly to Counterfactuals.

“PN works similarly to Counterfactuals.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

LIME only covers a local region and may not be generalisable.

“However, LIME only covers a local region and may not be generalisable.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

LIME is an optimal alternative for few models such as kNN in terms of computation cost.

“LIME is an optimal alternative for few models such as knn model in terms of computation cost.”

Probably not a substantial claim: may not pass S3 (load-bearing).

DiCE provides what-if explanations for model predictions.

“DiCE [45] is a package that gives the “what-if” explanations for the model prediction.”

• cites [45]Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

It is difficult to find the average in ICE plots.

“Third, it is difficult to find the average in ICE plots.”

Probably not a substantial claim: may not pass S3 (load-bearing).

DiCE implements counterfactual explanations that include perturbed features leading to the required result.

“It implements counterfactual explanations that include perturbed features which in turn will lead to the required result.”

• cites [45]Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Using CEM, the accuracy of an ML model can be enhanced by looking at mis-classified instances and using CEM explanations.

“Using CEM, the accuracy of the ML model can be enhanced by looking at cases of mis-classified instances and using the explanations provided by CEM.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Alibi comprises different types of explainers depending on the data.

“This library comprises different types of explainers depending on data we are dealing with.”

• cites [34]Not checked yet

  Not checked in the free run.

May combine two claims.

Counterfactual explainer methods work on black box models and are best suited for binary datasets.

“Counterfactual explainer method work on black box models and best suited for binary datasets.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

PP works similarly to Anchors.

“PP works similarly to Anchors.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Integrated Gradients is a local explanation method that explains individual predictions.

“It is a local method that helps explain each individual prediction.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

SHAP is computationally expensive on certain models such as KNN but runs fast on trees such as gradient boosted trees from XGBoost.

“SHAP is computationally expensive on certain models such as KNN but runs fast on trees, such as gradient boosted trees from XGBoost.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

It is computationally expensive to calculate Shapley Value compared to LIME.

“However, it is computationally expensive to calculate the Shapley Value in comparison to LIME.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

Counterfactuals guided by prototypes is much faster than counterfactual because prototypes speed up the search process by directing counterfactual to the prototype of a particular class.

“It is much faster than counterfactual, as prototype speed up the search process significantly by directing counterfactual to the prototype of a particular class.”

May combine two claims.

For datasets requiring complicated, non-interpretable models, there would be non-linearity at local regions, making LIME unable to deploy in such settings.

“There would be non-linearity at local regions for datasets requiring complicated, non-interpretable models.The inability to deploy LIME in such settings is a major pitfall.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

SHAP provides fast, reliable evaluations and incorporates TreeExplainer, which optimally estimates Shapley Values through XGBoost.

“Contrary, SHAP provides fast, reliable evaluations and incorporates TreeExplainer, which optimally estimates Shapley Values through XGBoost.”

Probably not a substantial claim: may not pass S3 (load-bearing).

At 3.8 cm Petal Length, the effects on each class are the same.

“From that, it can be concluded that the effects of the Petal Length on each class will be one and the same at 3.8 cm.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

LIME cannot handle XGBoost's requirement to use xgb.DMatrix() on input data.

“For example, LIME cannot handle the requirement of XGBoost to use xgb.DMatrix() on the input data.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

If the same perturbation space is allotted to both LIME and Anchors, Anchors will build a more valid region of instances that better describe the model's behaviour.

“If the same perturbation space is allotted to both LIME and Anchors, the latter will build a more valid region of instances which better describe the model's behaviour.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Alibi is an open-source Python library for ML-based model inspection and interpretation.

“Alibi [34] is an open source Python library which aims at ML-based model inspection and interpretation.”

• cites [34]Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

The Integrated Gradients technique, also called Path-Integrated Gradients, assigns an importance score to each input feature by using gradients of the model output.

“Integrated Gradients, also known as Path-Integrated Gradients is an XAI technique that assign an importance value to each feature of the input using the gradients of the model output [57].”

• cites [57]Not checked yet

  Not checked in the free run.

May combine two claims.

All three methods PDP, ICE, and Feature Importance consider features as independent entities, so if features are correlated, they create unlikely data points.

“In addition, all three methods consider the features as independent entity.Hence, if features are correlated, it will result in the creation of unlikely data points.”

May combine two claims.

PDP demonstrates global effects while concealing heterogeneous effects.

“PDP demonstrates global effects, concealing the heterogeneous effects.”

Probably not a substantial claim: may not pass S3 (load-bearing).

In the heart disease global SHAP example, higher values of ca indicate more chances of heart disease.

“High value of ca indicates more chances of Heart Disease.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

In the shirt example, the middle of the shirt can be mistaken for jeans or trousers, and this ambiguity makes those pixels negative attributions for the prediction.

“However, the middle portion of the shirt can be mistaken to be a part of a pair of jeans or trousers.Therefore, due to this ambiguity, they are the negative attributions for the prediction.”

Probably not a substantial claim: may not pass S3 (load-bearing).

A model produces a correct prediction when its positive attributions exceed its negative ones.

“All in all, we can affirm that when the positive attributions outweigh the negative attributions, the model makes the correct prediction.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

Counterfactual methods can be applied to classification with more than 3 target values, but performance degrades compared to binary classification.

“They can also be applied on classification with more than 3 target values, but the performance degrades as compared to binary classification.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Traditional variable importance algorithms provide results for the entire population, ignoring individual scenarios.

“The traditional variable importance algorithms provide results for the entire population ignoring individual scenarios.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Integrated Gradients provides feature attributions that can be positive or negative.

“This method provides the specific attributions of the feature inputs that are positive attributions and negative attributions.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Positive attributions are feature contributions that push a model toward a decision, while negative attributions push it away from a decision.

“Positive attributions are attributions that contribute or influence a model to make the decision whereas negative attributions are attributions that contribute or influence a model against a decision made.”

Probably not a substantial claim: may not pass S3 (load-bearing).

How much explanation is needed varies across experiments.

“The extent of requisite explainability differs from one experiment to another.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Knowing how much data an AI system uses and where it came from supports explainability.

“Being aware of data quantity and its provenance in AI systems ensures system explainability.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

XAI methods come in many forms, and each has advantages and disadvantages.

“XAI technologies utilize various schemes and approaches, each with its own pros and cons.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

A customer receiving shopping recommendations may not seek reasons for the suggested items.

“In the example of shopping recommendations, the customer may not want an explanation for the items recommended.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Figure 21's lower-left area shows the most successful AI applications, with low human involvement and low potential cost.

“The bottom left quadrant represents the most successful AI use cases, where both potential cost and human participation are low.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Models such as simple linear or tree-based ones can readily account for their decisions, yielding a particular insight or prediction.

“For example, few simple linear or tree-based models can easily explain the model decisions that lead to a specific insight or prediction.”

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based).

Designing a system frequently involves balancing requirements that conflict.

“System design often needs to balance competing requirements.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

The survey also considered distributed AI, where data are stored across several nodes in practical AI deployments, along with privacy, transparency, and fairness.

“We have explored our discussions beyond what has been gone so far in the XAI domain toward the idea of distributed AI, a paradigm that imposes storage of data on multiple nodes when implementing AI models in practice, including privacy, transparency, and fairness.”

Probably not a substantial claim: may not pass S3 (load-bearing).

The upper-right area of Figure 21 includes decisions like medical diagnosis and credit risk profiling, whose costs rise sharply.

“Decisions listed in the top right quadrant, e.g. for credit risk profiling, medical diagnosis etc, represent exponentially increasing cost.”

Probably not a substantial claim: may not pass S3 (load-bearing).

When designing an AI system, explainability is not always the main concern.

“Explainability may not always be the priority while designing an AI system.”

Classified as probably not a substantial claim. May combine two claims.

Ensemble and deep learning models, though often highly accurate, are treated as black boxes because their decisions are hard to evaluate.

“Conversely, complex ensemble and deep learning models often produce superlative performance, but they are considered as black-box models as it is very difficult to assess model's decisions.”

Probably not a substantial claim: may not pass S3 (load-bearing).

In Figure 21's upper-left area, mistakes can have catastrophic outcomes.

“The top left quadrant features decisions where any errors can result in disastrous consequences.”

Probably not a substantial claim: may not pass S3 (load-bearing).

AI systems in healthcare and finance rely on private data to decide and recommend.

“For instance, in healthcare and finance applications, AI systems are privy to confidential data for making decisions and offering recommendations.”

Probably not a substantial claim: may not pass S2 (checkable).

When machine learning informs critical decisions, explanation becomes essential.

“On the other hand, in situations where ML model is utilized for making crucial decisions, explainability is paramount.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

When XAI methods are used in different applications, interpretability, accuracy, and privacy vary.

“These methods when used in different applications, the interpretability, accuracy, and privacy varies.”