Abstract
An important task of information retrieval is to induce classifiers capable of categorizing text documents. The fact that the same document can simultaneously belong to two or more categories is referred by the term multi-label classification (or categorization). Domains of this kind have been encountered in diverse fields even outside information retrieval. This dissertation discusses one challenging aspect of text categorization: the documents (i.e., training examples) are characterized by an extremely large number of features. As a result, many existing machine learning techniques are in such domains prohibitively expensive. This dissertation seeks to reduce these costs significantly. The proposed scheme consists of two steps. The first runs a so-called baseline induction algorithm (BIA) separately on different versions of the data, each time inducing a different subclassifier---more specifically, BIA is run always on the same training documents that are each time described by a different subset of the features. The second step then combines the subclassifiers by a fusion algorithm: when a document is to be classified, each subclassifier outputs a set of class labels accompanied by its confidence in these labels; these outputs are then combined into a single multi-label recommendation. The dissertation investigates a few alternative fusion techniques, including an original one, inspired by the Dempster-Shafer Theory. The main contribution is a mechanism for assigning the mass function to individual labels from subclassifiers. The system's behavior is illustrated on two real-world data sets. As indicated, in each of them the examples are described by thousands of features, and each example is labeled with a subset of classes. Experimental evidence indicates that the method can scale up well and achieves impressive computational savings in exchange for only a modest loss in the classification performance. The fusion method proposed is also shown to be more accurate than other more traditional fusion mechanisms. For a very large multi-label data set, the proposed mechanism not only speeds up the total induction time, but also facilitates the execution of the task on a small computer. The fact that subclassifiers can be constructed independently and more conveniently from small subsets of features provides an avenue for parallel processing that might offer further increase in computational efficiency.