الفهرس | Only 14 pages are availabe for public view |
Abstract Label dependencies are the biggest influencing factor on performance multi-label classification, directly and indirectly. A point, which clearly distinguishes multi-label from multi-class problems, is the possible dependencies between the classes. Hence, the key challenge in multilabel learning is how to exploit this dependency effectively, because both empirical and theoretical studies have proved that it increases the performance of the learning process. Usually label correlations are given in advance. However, in some problem domains unfortunately this knowledge is unavailable, therefore new techniques are required to obtain the correlations between labels. We’ll try to exploit these correlations in this research. This is to be reached by discovering the correlation between classes using association rule, and then using a divide and conquer technique, the problem is divided into multiple smaller problems each sub-problem containing related labels is solved separately and then the result is integrated forming a global solution for the original problem. The thesis comprises seven chapters, these are organized as follows: In chapter two, we introduce concepts, notations and corresponding basic formal definitions required throughout this work. Furthermore, it discusses and review relevant existing and newly introduced evaluation measures, and provides an in depth study of multi-label data In chapter three, presents the three categories of methods for multi-label learning and discusses the advantages and disadvantages of each method, and then evaluate multi-label methods. |