Machine learning in software defect prediction in recent years has had a major impact on software engineering as a discipline. A major problem in this area is the need for a large amount of labeled data that machine learning models will use for learning in order to be able to predict software faults well. The motivation behind this paper comes from the ability to train models over a small amount of known, labeled data. The paper presents a new methodology for using a machine learning model that uses labeled and unlabeled data in combination with a genetic algorithm and an optimization cycle. The purpose of the genetic algorithm is to predict the values of dependent variables, that is, the presence of defects in software products. The idea of the paper is to use a small amount of labeled data for successful model learning, which will with the help of a genetic algorithm be enough to achieve satisfactory results. The idea proved promising and in some cases from three open source projects it achieved better results compared to the standard supervised learning model. However, to obtain more general conclusions, it would be necessary to conduct more extensive research that would include a larger amount of data from diverse projects.