An evolutionary approach for automatically extracting intelligible classification rules

Cutoff Threshold Decisions for Classification Algorithms with Risk Aversion

Classification algorithms predict the class membership of an unknown record. Methods such as logistic regression or the naïve Bayes algorithm produce a score related to the likelihood that a record belongs to a particular class. A cutoff threshold is then defined to delineate the prediction of one class over another. This paper derives analytic results for the selection of an optimal cutoff threshold for a classification algorithm that is used to inform a two-action decision in the cases of risk aversion and risk neutrality. The results provide insight to how the optimal cutoff thresholds relate to the associated costs and the sensitivity and specificity of the algorithm for both the risk neutral and risk averse decision makers. The optimal risk averse threshold is not reliably above or below the optimal risk neutral threshold, but the relation depends on the parameters of a particular application. The results further show the risk averse optimal threshold is insensitive to the size of the data set or the magnitude of the costs, but instead is sensitive to the proportion of positive records in the data and the ratio of costs. Numeric examples and sensitivity analysis derive further insight. Results show the percent value gap from a misspecified risk attitude increases as the specificity of the classification algorithm decreases.

Cartesian genetic programming for diagnosis of Parkinson disease through handwriting analysis: Performance vs. interpretability issues

In the last decades, early disease identification through non-invasive and automatic methodologies has gathered increasing interest from the scientific community. Among others, Parkinson's disease (PD) has received special attention in that it is a severe and progressive neuro-degenerative disease. As a consequence, early diagnosis would provide more effective and prompt care strategies, that cloud successfully influence patients' life expectancy. However, the most performing systems implement the so called black-box approach, which do not provide explicit rules to reach a decision. This lack of interpretability, has hampered the acceptance of those systems by clinicians and their deployment on the field. In this context, we perform a thorough comparison of different machine learning (ML) techniques, whose classification results are characterized by different levels of interpretability. Such techniques were applied for automatically identify PD patients through the analysis of handwriting and drawing samples. Results analysis shows that white-box approaches, such as Cartesian Genetic Programming and Decision Tree, allow to reach a twofold goal: support the diagnosis of PD and obtain explicit classification models, on which only a subset of features (related to specific tasks) were identified and exploited for classification. Obtained classification models provide important insights for the design of non-invasive, inexpensive and easy to administer diagnostic protocols. Comparison of different ML approaches (in terms of both accuracy and interpretability) has been performed on the features extracted from the handwriting and drawing samples included in the publicly available PaHaW and NewHandPD datasets. The experimental findings show that the Cartesian Genetic Programming outperforms the white-box methods in accuracy and the black-box ones in interpretability.