Нови методи за градење на хиерархиски повеќецелни класификатори

Abstract

Multi-label learning has received significant attention in the research community over the past few years, motivated by an increasing number of new applications. The latter includes semantic annotation of images and video (news clips, movies clips), functional genomics (gene and protein function), music categorization into emotions, text classification (news articles, web pages, patents, e-mails, bookmarks …), directed marketing and others. This has resulted in the development of a variety of multi-label learning methods. Previous research efforts have resulted in various methods and algorithms for solving multi-label and hierarchical multilabel classification and ranking problems. However, classification architecture with equal performance over various classification challenges has not been developed yet. Within this thesis an attempt is made to identify the advanced problems and challenges that the researchers face in the area of multi-label and hierarchical multi-label classification and ranking. The research undertaken in the preparation of this thesis resulted with three new algorithms for multi-label classification and ranking (Two stage architecture, Hybrid Decision Tree Architecture utilizing Local SVMs and Random forest of HOMER hierarchy predictive clustering trees). They are characterized with efficiency, scalability and high predictive performance. To asses the predictive performance of the proposed methods, an extensive experimental comparison with 12 multi-label learning methods using 18 evaluation measures over 11 benchmark datasets was made. The competing methods were selected based on their previous usage by the research community, the representation of different groups of methods and the variety of basic underlying machine learning methods. The obtained experimental results showed excellent performance regarding the training and classification speeds and high predictive performance in the recognition process.