Creating decision models for risk assessment of ski injuries is a challenging task. Ski injuries are rare events, but they carry a high cost, that is, can cause working or movement disabilities. Usually, ski risk assessment is performed on small-scale, case-controlled studies where the effect of a single factor is evaluated. Recently, data mining and machine learning algorithms are being employed for ski risk assessment and injury prediction. However, these models do not generally satisfy the need for interpretation of the decision model, do not provide explanations for the predictions, and in general do not ensure the completeness and consistency of decision rules. To make data mining and machine learning models useful, one needs to implement the aforementioned properties. Decision support systems are expected to have these properties; however, the process of building such decision support systems is still tedious - it has to consider human biases, assumptions, and subjective values, as well as focus on the decision problem being solved. We propose a method for extraction of decision models from data at hand. Our method DIDEX, Data Induced DEcision eXpert, builds models that have desirable properties for inclusion in decision support systems. The proposed method is used to build a decision model for ski injury prediction based on data from Mt. Kopaonik ski resort, Serbia. The results show that DIDEX generates up to a five times simpler model compared to the existing domain expert DEX models while having a 6% better predictive accuracy. Additionally, its predictive accuracy is comparable to similar machine learning algorithms, such as decision tree classifiers, random forest, and logistic regression.

With growing awareness of societal impact of artificial intelligence, fairness has become an important aspect of machine learning algorithms. The issue is that human biases towards certain groups of population, defined by sensitive features like race and gender, are introduced to the training data through data collection and labeling. Two important directions of fairness ensuring research have focused on (i) instance weighting in order to decrease the impact of more biased instances and (ii) adversarial training in order to construct data representations informative of the target variable, but uninformative of the sensitive attributes. In this paper we propose a Fair Adversarial Instance Re-weighting (FAIR) method, which uses adversarial training to learn instance weighting function that ensures fair predictions. Merging the two paradigms, it inherits desirable properties from both – interpretability of reweighting and end-to-end trainability of adversarial training. We propose four different variants of the method and, among other things, demonstrate how the method can be cast in a fully probabilistic framework. Additionally, theoretical analysis of FAIR models’ properties have been studied extensively. We compare FAIR models to 7 other related and state-of-the-art models and demonstrate that FAIR is able to achieve a better trade-off between accuracy and unfairness. To the best of our knowledge, this is the first model that merges reweighting and adversarial approaches by means of a weighting function that can provide interpretable information about fairness of individual instances.

Achieving fairness in algorithmic decision-making tools is an issue constantly gaining in need and popularity. Today, unfair decisions made by such tools can even be subject to legal consequences. We propose a new constraint that integrates fairness into data envelopment analysis (DEA). This allows calculating relative eficiency scores of decision-making units (DMUs) with fairness included. The proposed fairness constraint does not allow disparate impact to occur in eficiency scores, and enables the creation of a single data envelopment analysis for both privileged and discriminated groups of DMUs simultaneously. We show that the proposed method (FairDEA) produces an interpretable model that was tested on a synthetic dataset and on a real-world example, namely, the ranking between hybrid and conventional car designs. We provide the interpretation of the FairDEA method by comparing it to the basic DEA and the balanced fairness and eficiency method (BFE DEA). Along with calculating the disparate impact of the method, we performed a Wilcoxon rank-sum test to inspect for fairness in rankings. The results show that the FairDEA method achieves similar eficiency scores as other methods, but does not have disparate impact. Statistical analysis indicates that differences in ranking between groups are not statistically different, which means that ranking is also fair. This method contributes both to development of data envelopment analysis, and the inclusion of fairness in eficiency analysis.