Publications

2022

1. ShapChains

   Shapley chains: Extending Shapley values to Classifier chains

   Célia Wafa Ayad, Thomas Bonnier, Benjamin Bosch, and 1 more author

   Oct 2022

   Abs Code

   In spite of increased attention on explainable machine learning models, explaining multi-output predictions has not yet been extensively addressed. Methods that use Shapley values to attribute feature contributions to the decision making are one of the most popular approaches to explain local individual and global predictions. By considering each output separately in multi-output tasks, these methods fail to provide complete feature explanations. We propose Shapley chains to overcome this issue by including label interdependencies in the explanation design process. Shapley chains assign Shapley values as feature importance scores in multi-output classification using classifier chains, by separating the direct and indirect influence of these feature scores. Compared to existing methods, this approach allows to attribute a more complete feature contribution to the predictions of multi-output classification tasks. We provide a mechanism to distribute the hidden contributions of the outputs with respect to a given chaining order of these outputs. Moreover, we show how our approach can reveal indirect feature contributions missed by existing approaches. Shapley chains help to emphasize the real learning factors in multi-output applications and allows a better understanding of the flow of information through output interdependencies in synthetic and real-world datasets.

2023

1. XAICancer

   Which Explanation Makes Sense? A Critical Evaluation of Local Explanations for Assessing Cervical Cancer Risk Factors

   Célia Wafa Ayad, Thomas Bonnier, Benjamin Bosch, and 2 more authors

   Aug 2023

   Abs Code

   Cervical cancer is a life-threatening disease and one of the most prevalent types of cancer affecting women worldwide. Being able to adequately identify and assess factors that elevate risk of cervical cancer is crucial for early detection and treatment. Advances in machine learning have produced new methods for predicting cervical cancer risk, however their complex black-box behaviour remains a key barrier to their adoption in clinical practice. Recently, there has been substantial rise in the development of local explainability techniques aimed at breaking down a model’s predictions for particular instances in terms of, for example, meaningful concepts, important features, decision tree or rule-based logic, among others. While these techniques can help users better understand key factors driving a model’s decisions in some situations, they may not always be consistent or provide faithful predictions, particularly in applications with heterogeneous outcomes. In this paper, we present a critical analysis of several existing local interpretability methods for explaining risk factors associated with cervical cancer. Our goal is to help clinicians who use AI to better understand which types of explanations to use in particular contexts. We present a framework for studying the quality of different explanations for cervical cancer risk and contextualise how different explanations might be appropriate for different patient scenarios through an empirical analysis. Finally, we provide practical advice for practitioners as to how to use different types of explanations for assessing and determining key factors driving cervical cancer risk.