ABSTRACT
In this paper we explore the reliability of contexts of machine learning (ML) models. There are several evaluation procedures commonly used to validate a model (precision, F1 Score and others); However, these procedures are not linked to the evaluation of learning itself, but only to the number of correct answers presented by the model. This characteristic makes it impossible to assess whether a model was able to learn through elements that make sense of the context in which it is inserted. Therefore, the model could achieves good results in the training stage but poor results when the model needs to be generalized. When there are many different models that achieve similar performance, the model that presented the highest number of hits in training does not mean that this model is the best. Therefore, we created a methodology based on Item Response Theory that allows us to identify whether an ML context is unreliable, providing an extra and different validation for ML models.
