The use of the Choquet integral as a classifier in binary data: Preliminary results in the detection of COVID-19

In several problems involving the aggregation of complex information, it is common to use functions that allow modeling interactions between attributes. In ranking problems, for example, one generally uses the Choquet integral to aggregate the criteria information. However, although the Choquet integral is widely used in multi-criteria decision problems, there are few studies that apply this function as a classifier. In this work, we consider its use as a classifier in scenarios where the dataset is binary. This situation is typical in medical data, since much information is of the type “has” or “does not have” a symptom. An interest in this scenario is that it results in a particular case of the Choquet integral, which makes it easier to adjust the parameters and simplify the transformation from them to the so-called Shapley values and interaction indices, which are widely used in machine learning to extract interpretation from models. Basically, these parameters indicate both the impact of each attribute on the classification and the synergistic or redundant effects between them. In order to illustrate the use of the Choquet integral as a classifier, we used a COVID-19 detection dataset. This binary dataset contains information on people who presented various symptoms (such as fever, cough, among others) and who took the COVID-19 test, with the result being positive or negative. Using the proposed model, we could understand which symptoms are most associated with the detection of this disease in the tested individuals.