REPOSITORY > RESULTS

Doctoral dissertation

Multi-View Canonical Correlation Analysis

Author(s): Jan Rupnik (Author), Dunja Mladenić (Supervisor), John Shawe-Taylor (Co-Supervisor)

Thesis defense date: 17.06.2016

Organization: MPŠ - Mednarodna podiplomska šola Jožefa Stefana

PID: 20.500.12556/ReVIS-15043

Views: 4 | Downloads: 0

Abstract

Matrix factorization represents a popular approach in pattern analysis and is used to
tackle many problems, such as: collaborative ltering, imputing missing data, denoising
data, dimensionality reduction, data visualization and exploratory analysis.
This thesis is focused on factorization based pattern analysis methods for multiview
learning problems: that is problems where each data instance is represented by multiple
views of an underlying object, encoded by multiple feature sets. As an example of a
multiview problem consider a dataset where each instance has two representations: a
visual image and a textual description. The patterns of interest are pairs of functions over
images and texts that are strongly related over the observed data.
Canonical Correlation Analysis (CCA) is designed to extract patterns from data sets
with two views. This thesis focuses on two generalizations of CCA, which were proposed in
the literature: Sum of Correlations (SUMCOR) and Sum of Squared Correlations SSCOR.
The SUMCOR problem formulation is interesting from the optimization perspective by its
own right, since it emerges in other problems as well.
We study several aspects of the generalizations. We rst present a provably convergent
novel algorithm for nding non-linear higher order patterns, which is based on an iterative
approach for solving multivariate eigenvalue problems. We show that SUMCOR in general
is NP-hard and then study its reformulation to a computationally tractable Semidenite
Programming (SDP) problem. Based on the reformulation we derive several computationally
feasible bounds on global optimality, which complement the locally optimal solutions.
We introduce a new preprocessing step for dealing with large scale SDP problems that
arise from an application to cross-lingual text analysis. We investigated how to apply our
methods to real datasets with missing data. The particular structure of missing data in the
problem considered leads to a simplication of the SSCOR optimization problem, which is
reduced to a tractable eigenvalue problem. We show how the algorithms apply to building
cross-lingual similarity models and apply the models on the task of cross-lingual cluster
linking. The approach to cross-lingual cluster linking is used in a real-time global analysis
of news streams in multiple languages.

Attachments

Cite this work