Événements du LITIS
- Cet évènement est passé
Séminaire » Dimensionality reduction: towards better scalability » par John Lee
09/15/2016 @ 14:00 - 16:00
Le jeudi 15 septembre se déroulera un séminaire sur le sujet suivant:
Dimensionality reduction: towards better scalability
Ce séminaire sera animé par John Lee de l’UC Louvain et se tiendra à partir de 14H dans la salle de la bibliothèque LITIS à l’université de Rouen, campus du Madrillet.
N.B. La présentation aura lieu en français et traitera en particulier de l’utilisation des méthodes de réduction de dimension dans le cadre de grandes masses de données.
» Dimensionality reduction (DR) aims at representing high-dimensional data in low-dimensional spaces, while preserving important structural properties, like for example (dis)similarities or neighbourhood relationships. The vast majority of DR methods work in a unsupervised way: they process data features without taking into account additional information like class labels, which are then sometimes used to assess DR quality. Dimensionality reduction can be used for different purposes, ranging from exploratory data analysis (visual inspection) to data compression before subsequent processing. In the latter case, DR can be seen as a way to defeat the so-called curse of dimensionality, which makes many complex analysis tasks like regression or classification much more difficult in high-dimensional spaces than in low-dimensional ones.
High dimensionality is not the only issue that analysts have to face in the current era of data plethora. Data collection and storage becomes easier and cheaper every day. Processing large amounts of data raises many issues, in terms of algorithmic complexity (time and memory consumption), workload distribution (vectorised, parallel, or distributed architectures), and efficient visual presentation of the results. Politics and media have coined the term ‘Big Data’ to refer to these problems and the effort to alleviate them.
This talk briefly revisits past and recent history of DR and shows how the issue of large data sets is currently being dealt with. First, we will sketch the various paradigms of DR and how quality assessment (QA) of this unsupervised task can be carried out. Next, we will show how DR methods and QA frameworks that were published in the recent years can be extended in order to deal with big data to some extent. Finally, we will summarise open questions and perspectives for the near future »
Venez nombreux !