责任导师 : Patrick GALLINARI

助理责任导师 : BASKIOTIS Nicolas

Representation Learning for large scale classification

The past decades have seen the rise of new technologies that simplify information sharing. Today, a large part of the data is accessible to a large number of users. In this thesis, we propose to study the problems of document annotations that lead to many applications such as information retrievals ones. We will be interested in the field of extreme classification which characterizes the task of automatic classification when the number of labels is important. Many difficulties arise from the size and complexity of considered data: the prediction time, storage and relevance of the prediction are the most representative. Recent research dealing with this issue is now based on three types of approaches: ensembling approaches learning a large set of simple classifiers; “hierarchical” methods organizing a structure of simple classifiers; approaches by representations plunging documents into small spaces. In this thesis, we will study the approaches of classification by representation. Through our contributions, we will propose different approaches to overcome the problems of prediction time and representation space structure. First, we will study discrete representations with the objective to find the best possible representations while ensuring a low inference time. In a second step, we will consider hyperbolic representations in order to take advantage of the qualities of this space for structured data.

2017-2020 刊物