责任导师 : Patrick GALLINARI

助理责任导师 : BASKIOTIS Nicolas

Etude de la classification dans un très grand nombre de catégories

The increase in volume of the data nowadays is at the origin of new problematics for which machine learning does not possess adapted answers. The usual classification task which requires to assign one or more classes to an example is extended to problems with thousands or even millions of different classes.
Those problems bring new research fields like the complexity reduction of the classification process. That classification process has a complexity usually linear with the number of classes of the problem, which can be an issue if the number of classes is too large.
Various ways to deal with those new problems have emerged like the construction of a hierarchy of classifiers or the adaptation of ECOC ensemble methods.
The work presented here describes two new methods to answer this extreme classification task.
The first one consists in a new asymmetrical measure to help the partitioning and the hierarchisation of the classes in order to build a classes tree.
The second one proposes a sequential way to aggregate effectively the most interesting classifiers.

2014-2020 刊物