2008 | OriginalPaper | Buchkapitel
A Diffusion Framework for Dimensionality Reduction
verfasst von : Alon Schclar
Erschienen in: Soft Computing for Knowledge Discovery and Data Mining
Verlag: Springer US
Aktivieren Sie unsere intelligente Suche, um passende Fachinhalte oder Patente zu finden.
Wählen Sie Textabschnitte aus um mit Künstlicher Intelligenz passenden Patente zu finden. powered by
Markieren Sie Textabschnitte, um KI-gestützt weitere passende Inhalte zu finden. powered by
Many fields of research deal with high-dimensional data sets. Hyperspectral images in remote sensing and in hyper-spectral microscopy, transactions in banking monitoring systems are just a few examples for this type of sets. Revealing the geometric structure of these data-sets as a preliminary step facilitates their efficient processing. Often, only a small number of parameters govern the structure of the data-set. This number is the true dimension of the data-set and is the motivation to reduce the dimensionality of the set. Dimensionality reduction algorithms try to discover the true dimension of a data set.
In this chapter, we describe a natural framework based on diffusion processes for the multi-scale analysis of high-dimensional data-sets (Coifman and Lafon, 2006). This scheme enables us to describe the geometric structures of such sets by utilizing the Newtonian paradigm according to which a global description of a system can be derived by the aggregation of local transitions. Specifically, a Markov process is used to describe a random walk on the data set. The spectral properties of the Markov matrix that is associated with this process are used to embed the data-set in a low-dimensional space. This scheme also facilitates the parametrization of a data-set when the high dimensional data-set is not accessible and only a pair-wise similarity matrix is at hand.