Efficient algorithms for extreme classification

01 October 2019 → 30 September 2021
Research Foundation - Flanders (FWO)
Research disciplines
No data available
Machine learning extreme classification large-scale learning
Project description

Recent years have witnessed an increasing interest in the use of machine learning techniques for categorizing information. Computers have demonstrated a recognition rate better than or comparable to human performance in several tasks, such as visual recognition and recommendation systems. Despite these success stories, an important gap continues to exist for a long time between the scalability and the complexity of the problems handled by humans and those solved by computers. Most of the problems solved by computers using machine learning techniques involve at most few hundreds of categories (labels) while humans are able to discriminate at least tens of thousands of categories. Statistical and computational challenges posed by the presence of an extremely large number of labels have opened a new line of research, which is referred to as extreme classification. In such a setting, standard methods that have a training and prediction cost linear in the number of labels become intractable. Algorithms for extreme classification often require a sublinear time and space complexity. Another major challenge lies in that fact that getting reliable training information is a difficult task, therefore, one has to deal with the problem of learning from unreliable information. Given these challenges, this project aims to develop new algorithms for efficiently solving extreme classification problems. The proposed algorithms will be supported by empirical as well as theoretical studies.