Recent Studies of XAI - Review

Abstract

Over the past years, there has been an increasing concern regarding the risk of bias and discrimination in algorithmic systems, which received significant attention amongst the research communities. To ensure the system's fairness, various methods and techniques have been developed to assess and mitigate potential biases. Such methods, also known as "Formal Fairness", look at various aspects of the system's advanced reasoning mechanism and outcomes, with techniques ranging from local explanations (at feature level) to visual explanations (saliency maps). Another aspect, equally important, represents the perception of the users regarding the system's fairness. Despite a decision system being provably "Fair", if the users find it difficult to understand how the decisions were made, they will refrain from trusting, accepting, and ultimately using the system altogether. This raised the issue of "Perceived Fairness" which looks at means to reassure users of a system's trustworthiness. In that sense, providing users with some form of explanation on why and how certain outcomes resulted, is highly relevant, especially nowadays as the reasoning mechanisms increase in complexity and computational power. Recent studies suggest a plethora of explanation types. The current work aims to review the recent progress in explaining systems' reasoning and outcome, categorize and present it as a reference for the state-of-the-art fairness-related explanations review. CCS CONCEPTS• Human-centered computing → Human computer interaction (HCI); • Computing methodologies → Artificial intelligence.

Explainable Artificial Intelligence (XAI): A Systematic Literature Review on Taxonomies and Applications in Finance

“…Works that support the category definition Stage [6], [8]- [11], [13], [15], [22], [23] Model [6], [8], [10], [12], [13], [15], [22], [23] Scope [6], [8], [10], [14], [15], [22] A "post-hoc" XAI method is named after the fact that it acts after predictions are made, not knowing how the predictor model made its decisions (e.g., LIME ( [24]). It is a surrogate model since it tries to simplify the function of the black-box model by sampling, perturbing data, and weighing the distance between instances to generate an approximation of the black-box model.…”

“…The final proposal for a category for XAI methods is Scope, intending to separate XAI methods on whether they are used to help understand the general behavior of the model, that is, if these techniques provide global interpretability or if they try to explain singular or a limited group of instances of data, that is, local interpretability [6]. This category is largely accepted within the reviewed literature, where it is found as a main category for classifying XAI methods [6], [8], [10], [14], [15], [22].…”

“…This category is a contender for relevancy when defining a taxonomy since other authors support this category, even if under different names [15]. Another work favoring the categorization of results is [14], although this category is named "Presentation Format," showing two sub-classes on whether the generated explanation is textual (when explanations are generated using natural language techniques) or visual, focusing on providing a visual explanation, for example, via graphs or images. We can also find "Result" among other categories mentioned in [23].…”

“…The definition of a taxonomy for XAI methods, primarily adapted from other papers and with several categories that include but are not limited to the domain of application of the method, is presented in [14]. After reviewing previous work, another paper focusing on the definition of a proper taxonomy of XAI methods concludes with a proposal for a taxonomy trying to adapt the taxonomies found in their review [15].…”

XAI-based Clinical Decision Support System: A Systematic Review

“…Complexity-based techniques make machine learning or deep learning models fully interpretable. Interpretability can be categorized into intrinsic interpretability [96] and post hoc interpretability [72] depending on the viewpoint. In general, intrinsic interpretability indicates that a model with a simple architecture can be explained by the trained model itself, whereas post hoc interpretability means that the trained model has a complex architecture and must be retrained to explain this phenomenon.…”