Black-box machine learning models are used in an increasing number of high-stakes domains, and this creates a growing need for Explainable AI (XAI). However, the use of XAI in machine learning introduces privacy risks, which currently remain largely unnoticed. Therefore, we explore the possibility of an explanation linkage attack , which can occur when deploying instance-based strategies to find counterfactual explanations. To counter such an attack, we propose k -anonymous counterfactual explanations and introduce pureness as a metric to evaluate the validity of these k -anonymous counterfactual explanations. Our results show that making the explanations, rather than the whole dataset, k -anonymous, is beneficial for the quality of the explanations.

Abstract The complexity of state-of-the-art modeling techniques for image classification impedes the ability to explain model predictions in an interpretable way. A counterfactual explanation highlights the parts of an image which, when removed, would change the predicted class. Both legal scholars and data scientists are increasingly turning to counterfactual explanations as these provide a high degree of human interpretability, reveal what minimal information needs to be changed in order to come to a different prediction and do not require the prediction model to be disclosed. Our literature review shows that existing counterfactual methods for image classification have strong requirements regarding access to the training data and the model internals, which often are unrealistic. Therefore, SEDC is introduced as a model-agnostic instance-level explanation method for image classification that does not need access to the training data. As image classification tasks are typically multiclass problems, an additional contribution is the introduction of the SEDC-T method that allows specifying a target counterfactual class. These methods are experimentally tested on ImageNet data, and with concrete examples, we illustrate how the resulting explanations can give insights in model decisions. Moreover, SEDC is benchmarked against existing model-agnostic explanation methods, demonstrating stability of results, computational efficiency and the counterfactual nature of the explanations.

Research on scalable machine learning algorithms has gained a considerable amount of traction since the exponential growth in data assets during the past decades. Many Big Data applications resort to somewhat "simple" data modelling techniques due to the computational constraints associated with more complex models. Simple models, while being very efficient to estimate, often fail to capture some of the finer details of more complex datasets. In this manuscript, we explore the idea that complex large scale classification can be tractable using a process of iterative refining. In such a process, we focus on non-linearities of the data only after having first found an approximate linear model. This knowledge is then incorporated into the nonlinear model implicitly, allowing the non-linear model to focus on important parts of the data after a rough first estimation. This in turn reduces overall training time and allows for a richer model representation, eventually leading to more predictive power.

Many of the state-of-the-art data mining techniques introduce nonlinearities in their models to cope with complex data relationships effectively. Although such techniques are consistently included among the top classification techniques in terms of predictive power, their lack of transparency renders them useless in any domain where comprehensibility is of importance. Rule-extraction algorithms remedy this by distilling comprehensible rule sets from complex models that explain how the classifications are made. This paper considers a new rule extraction technique, based on active learning. The technique generates artificial data points around training data with low confidence in the output score, after which these are labeled by the black-box model. The main novelty of the proposed method is that it uses a pedagogical approach without making any architectural assumptions of the underlying model. It can therefore be applied to any black-box technique. Furthermore, it can generate any rule format, depending on the chosen underlying rule induction technique. In a large-scale empirical study, we demonstrate the validity of our technique to extract trees and rules from artificial neural networks, support vector machines, and random forests, on 25 data sets of varying size and dimensionality. Our results show that not only do the generated rules explain the black-box models well (thereby facilitating the acceptance of such models), the proposed algorithm also performs significantly better than traditional rule induction techniques in terms of accuracy as well as fidelity.

With the globalisation of the world's economies and ever-evolving financial structures, fraud has become one of the main dissipaters of government wealth and perhaps even a major contributor in the slowing down of economies in general. Although corporate residence fraud is known to be a major factor, data availability and high sensitivity have caused this domain to be largely untouched by academia. The current Belgian government has pledged to tackle this issue at large by using a variety of in-house approaches and cooperations with institutions such as academia, the ultimate goal being a fair and efficient taxation system. This is the first data mining application specifically aimed at finding corporate residence fraud, where we show the predictive value of using both structured and fine-grained invoicing data. We further describe the problems involved in building such a fraud detection system, which are mainly data-related (e.g. data asymmetry, quality, volume, variety and velocity) and deployment-related (e.g. the need for explanations of the predictions made).

The paper provides a comment from the authors regarding the correlation as a heuristic for accurate and comprehensible ant colony optimization-based classifiers. Baig et al. proposed a new classification rule mining algorithm in IEEE TEC. The technique introduces a correlation-based function for the ant colony optimization (ACO) component. The authors compare variants of the technique with several existing ACO-based algorithms and the state-of-the art RIPPER algorithm. The AntMiner+ algorithm, proposed by Martens et al. (2007) in IEEE TEC, was also included in the comparison. However, the reported performance is far below what is expected and is more akin to making random predictions.

A predictive model is required to be accurate and comprehensible in order to inspire confidence in a business setting. Both aspects have been assessed in a software effort estimation setting by previous studies. However, no univocal conclusion as to which technique is the most suited has been reached. This study addresses this issue by reporting on the results of a large scale benchmarking study. Different types of techniques are under consideration, including techniques inducing tree/rule-based models like M5 and CART, linear models such as various types of linear regression, nonlinear models (MARS, multilayered perceptron neural networks, radial basis function networks, and least squares support vector machines), and estimation techniques that do not explicitly induce a model (e.g., a case-based reasoning approach). Furthermore, the aspect of feature subset selection by using a generic backward input selection wrapper is investigated. The results are subjected to rigorous statistical testing and indicate that ordinary least squares regression in combination with a logarithmic transformation performs best. Another key finding is that by selecting a subset of highly predictive attributes such as project size, development, and environment related attributes, typically a significant increase in estimation accuracy can be obtained.

The five papers in this special section focus on white-box nonlinear prediction models.

Neural networks are often selected as tool for software effort prediction because of their capability to approximate any continuous function with arbitrary accuracy. A major drawback of neural networks is the complex mapping between inputs and output, which is not easily understood by a user. This paper describes a rule extraction technique that derives a set of comprehensible IF-THEN rules from a trained neural network applied to the domain of software effort prediction. The suitability of this technique is tested on the ISBSG R11 data set by a comparison with linear regression, radial basis function networks, and CART. It is found that the most accurate results are obtained by CART, though the large number of rules limits comprehensibility. Considering comprehensible models only, the concise set of extracted rules outperform the pruned CART tree, making neural network rule extraction the most suitable technique for software effort prediction when comprehensibility is important.

Support vector machines (SVMs) are currently state-of-the-art for the classification task and, generally speaking, exhibit good predictive performance due to their ability to model nonlinearities. However, their strength is also their main weakness, as the generated nonlinear models are typically regarded as incomprehensible black-box models. In this paper, we propose a new Active Learning-Based Approach (ALBA) to extract comprehensible rules from opaque SVM models. Through rule extraction, some insight is provided into the logics of the SVM model. ALBA extracts rules from the trained SVM model by explicitly making use of key concepts of the SVM: the support vectors, and the observation that these are typically close to the decision boundary. Active learning implies the focus on apparent problem areas, which for rule induction techniques are the regions close to the SVM decision boundary where most of the noise is found. By generating extra data close to these support vectors that are provided with a class label by the trained SVM model, rule induction techniques are better able to discover suitable discrimination rules. This performance increase, both in terms of predictive accuracy as comprehensibility, is confirmed in our experiments where we apply ALBA on several publicly available data sets.

Ant colony optimization (ACO) can be applied to the data mining field to extract rule-based classifiers. The aim of this paper is twofold. On the one hand, we provide an overview of previous ant-based approaches to the classification task and compare them with state-of-the-art classification techniques, such as C4.5, RIPPER, and support vector machines in a benchmark study. On the other hand, a new ant-based classification technique is proposed, named AntMiner+. The key differences between the proposed AntMiner+ and previous AntMiner versions are the usage of the better performing MAX-MIN ant system, a clearly defined and augmented environment for the ants to walk through, with the inclusion of the class variable to handle multiclass problems, and the ability to include interval rules in the rule list. Furthermore, the commonly encountered problem in ACO of setting system parameters is dealt with in an automated, dynamic manner. Our benchmarking experiments show an AntMiner+ accuracy that is superior to that obtained by the other AntMiner versions, and competitive or better than the results achieved by the compared classification techniques.