Dimensionality reduction for clustering with deep neural networks

Abstract

[eng] Nowadays, high dimensional data is ubiquitous: you can think for example in images, videos or texts. Unfortunately, this property can harm seriously the performance of some algorithms. In this project, I analyse how dimensionality reduction can help clustering improve its performance. In order to do that, I distinguish three di erent clustering strategies: Traditional, two-stages and deep clustering. In the rst one, the clustering is applied to the raw data while in the other two it is applied to a low dimensional representation. I focus especially on the latter approach, which has shown promising performance in the last years. The di erences between these approaches are illustrated doing a series of experiments and visualisations and comparing the results.

[cat] Actualment, les dades d'alta dimensi o s on omnipresents: es pot pensar, per exemple, en imatges, v deos o textos. Malauradament, aquesta propietat pot perjudicar greument el rendiment d'alguns algorismes. En aquest projecte, analitzo com la reducci o de dimensionalitat pot ajudar a que el clustering millori el seu rendiment. Per fer-ho, distingeixo tres estrat egies de clusteritzaci o diferents: la tradicional, la de dues etapes i el Deep clustering. En la primera, l'agrupaci o s'aplica a les dades brutes mentre que en les altres dos s'aplica a una representaci o de baixa dimensi o. En el treball em centro especialment en aquest ultim enfocament, que ha demostrat un rendiment prometedor en els darrers anys. Les difer encies entre aquestes estrat egies s'il·lustren fent una s erie d'experiments i visualitzacions i comparant els resultats.