Machine learning (ML) models are victims of the oracle problem, i.e. it is not possible to know with absolute confidence an output for a given input. This prevents them from being evaluated using conventional software engineering techniques. However, there is an approach called “metamorphic relation” that reduces the oracle problem and helps to evaluate ML models. Unlike conventional tests, a metamorphic relation does not check if an input produces a specific output, but checks if a relationship between inputs and outputs is respected. Naturally, metamorphic relations have already been proposed in the literature, either to evaluate the behavior of a specific ML model, or to evaluate the general behavior of any ML model. The purpose of this paper is to propose new metamorphic relations to complement those of the literature, in order to propose a more complete methodology for evaluating ML models. So, in order to challenge this methodology, all these metamorphic relations are used to evaluate 21 different machine learning algorithms.