The Problem of First Story Detection in Multiaspect Text Categorization

The new concept of multiaspect text categorization (MTC), recently introduced in a series of our papers, may be viewed as a combination of the classic and well-known text categorization (TC) and some kind of sequential data classification. The first aspect of the problem, i.e., the assignment of a document to a category, may be addressed using one of the well-known techniques such as, e.g., the k-nearest neighbors method. The second aspect is, however, less standard and boils down to the assignment of a document to one of the sequences, called cases, of documents maintained within a category. Cases cannot be treated in the same way as categories as, first, they contain an ordered—by the time of arrival—set of documents, and second, they are usually represented in a training dataset by a (relatively) small number of documents. Moreover, it is assumed that new cases can emerge during the document collection lifetime. Hence, the assignment of a document to a case is a challenging task by itself, and then the deciding if a document starts a new case is even more difficult. In this paper, we deal with the latter problem, discussing it in the broader perspective of sequential data mining and comparing a number of approaches to solve it.