Chance constrained uncertain classification via robust optimization (Q633103)

The classification issue is one of the most common, but, at the same time, one of the most fruitful domains in the machine learning methodology. Ideally, the data used in the classification process have to be certain and neat. Unfortunately, this is not the case encountered in real-world problems. In real applications, data with different degree of uncertainty are met and there is the need to deal with such data in an appropriate way. This paper analyses the problem of building an efficient method to deal with uncertain data by using Bernstein bounding schemes (BBS). The main idea is to pose the uncertain data classification problem as a convex chance-constrained program and relax it as a second order cone program using BBS. Thus, various robust formulations employing different levels of partial information have been derived. The paper also provides numerical experiments which compare this new approach with existing methodologies in order to highlight its effectiveness.