Automated Learning Approaches

Outliers in a set of data are elements which are anomalous with respect to the majority of the data in the set. When collecting large amount of data, the problem of outliers detection arises. In order to solve this problem, a set of separating rules are used. Rules can be either given by a human expert or automatically generated. In the latter case, rules can be learned by the available data by using a data training set, i.e. data whose classification is already known. This talk is concerned with the problem of automatic detection of outliers when a training set is given. Several approaches to this problem are proposed in literature. In particular, we will briefly discuss the following particularly interesting approaches. Emphasis will be expecially given to the discrete ones. We begin with the Logical Analysis of Data (Boros-Hammer-Ibaraki-Kogan), which solves data classification problems by min-cover formulations. Subsequenely, the approach of Support Vector Machines (Vapnik) is described. After this, a brief overview on rule identification by their implementation as Neural Networks (Caianiello, Fletcher-Hinde) is presented. Then, concept lattice (Birkhoff) oriented techniques are illustred. Finally, a recent approach using a min-cut formulation over a graph representing our data (Blum-Chawla) is described. Moreover, some new ideas and problems concerning detection of incosistences and redundancies in the set of rules are discussed.