Dimensionality estimation methods and dimensionality reduction techniques applied to real biological time series data

Talk by Jérôme Lapuyade-Lahorgue

Abstract:

Determining the number of components in dimensionality reduction techniques constitutes a preliminary treatment to dimensionality reduction. Dimensionality reduction is divided into two important

parts:  the Factorial Analysis (FA) and the Low-Dimensional Representation (LDR). The FA includes all techniques which allows to estimate a reduced number of factors able to represent the data as conformly as possible. This number of factors appears to be the intrinsic dimension of the data. In the second hand, the LDR aims to find a lower dimensional space which preserves some properties of the observed data. In both methods, the intrinsic dimension is supposed to be known. However, this dimension is not known in general. In this talk, we will present briefly different methods to estimate this dimensionality in linear and in no linear case. Finally, we will test two particular methods on real biological data, these two particular methods are based on the homogeneity of the measure according to the dimension.

Host: Kim Steenstrup Pedersen