Matrix Methods in Data Mining and Pattern Recognition

1
Matrix Methods in Data Mining and Pattern
Recognition Prof. Lars Eldén (University of
Linköping) June 8 12, 2009 ITMAN Research
School, DTU Informatics, Lyngby, Denmark
Course description The PhD course covers several
very powerful numerical linear algebra techniques
for solving problems in different areas of data
mining and pattern recognition, with focus on the
applications of the methods, and illustrated with
Matlab examples. The goal is to give the student
a set of tools that may be tried as they are,
and also can be modified to be useful for
particular applications. Course requirements
The participants must have a fundamental
knowledge of numerical analysis and linear
algebra and must be able to program in Matlab.
Day 1 Introduction Matrices in data mining and
pattern recognition, Orthogonal transformations,
least squares, QR decomposition, Singular Value
Decomposition (SVD). Day 2 Classification of
handwritten digits Fundamental subspaces of a
matrix, low rank approximation by SVD.
Classification using SVD bases. Tangent distance.
Day 3 Text Mining The vector space model,
Latent Semantic Indexing, clustering methods.
Day 4 Pagerank Google pagerank web link graph,
random surfer model, computation of the
eigenvector of the link graph matrix.
Non-standard applications of pagerank. HITS
(Hypertext Induced Topic Search). Day 5
Advanced Matrix Methods Nonnegative matrix
factorization, Krylov subspace methods.
Applications in text mining and classification.
The course consists of lectures and computer
exercises, running each day from 9 to 4.
Course evaluation Final evaluation will be based
on a report which is agreed upon during
the course and which corresponds to about 30
hours of workload. The report must be handed in
no less than 4 weeks after completion of course.
Registration and deadline Please sign up with
Lone Hegelund, IMM lhe_at_imm.dtu.dk, no later than
XXX. DTU Informatics www.imm.dtu.dk.