Dipartimento di Scienze Statistiche

PhD program in Statistics DSS Statistics Seminar Salvatore Ingrassia Dipartimento di Economia e Impresa, Università di Catania.   In the statistical learning framework, we assume that the learning set is an unbiased sample of the population of interest and the test set is used to assess the fitted model. Anyway, in real-world applications, the available data for the learning set can result to be a biased sample of the target population and therefore a classifier trained on these data may result to be quite inadequate. In other words, there is a “shift" between the two domains. There are many reasons that can explain this dissimilarity: for instance, collecting new labeled data is often time-consuming, costly, or even infeasible, especially when the statistical properties of the population evolve over time. In some other cases, we are provided with a large amount of unlabeled data, but only a small amount of labeled data. In the machine learning literature, the area dealing with this kind of problem is called Domain Adaptation. In this seminar, we present main ideas of the Domain adaptation in a statistical setting. Numerical studies will be further presented based on both simulated and real datasets.