Florent Nicart - Université de Rouen (Pattern
Discovery in Strings and Sequences, application to patterns in text)
Pattern Discovery vs Pattern Matching; Memes and n-grams; Indexing
sequences; Surprising sequences; Web Mining; Suffix trees
Jean Philippe Vert - Mines Paris Tech [Pattern Analysis over Graphs, and Bioinformatics Applications]
1. Classification and regression over graphs. Overview:
positive definite graph kernels based on walk, subtrees etc.., as well
as other non p.d. similarity functions (eg from graph matching) that
can be used to compare graphs and do classification/regression with
kernel methods. Applications: QSAR in chemistry, image classification
2. Detecting patterns in the context of regression or classification with a graph as prior knowledge over the features.
in a classical regression/classification problem over high-dimensional
vectors. Control the complexity, by using priors that can be derived
from the graph over the vectors, and how they can be used as penalty
functions for classification and regression. This will cover diffusion
kernels and other kernels over graphs, fused lasso, structured group
lasso. Application in bioinformatics.
John Shawe-Taylor - University College London (Statistical significance and stability analysis for patterns)
Nello Cristianini - University of Bristol (Principles of Pattern Analysis, and their applications in Science and Industry)
Bart Goethals - University of Antwerp - Frequent Pattern Mining - Apriori, Eclat, Fp-growth, and a few more set mining algorithms
and some general optimization tricks - Closed set mining techniques -
Maximal set mining - Non-derivable set mining - mining sets in streams
Elisa Ricci, Idiap Patterns in vector spaces. *Eigenvalue problems
- finding relations between two datasets: Principal component analysis (PCA) and canonical correlation analysis (CCA)
- regularized PCA/CCA - kernel PCA/CCA *Least squares problems
-linear regression (LSR/RR) and Fisher's discriminant analysis (FDA)
*Support vector machines -primal/dual QP problems
-linear/non linear SVMs
-practical issues (parameters/kernel selection, solving QP problems)
-multiclass classification and learning with structured output "
Fabio Roli - Universita' di Cagliari - Pattern Recognition with Multiple Classifier Systems
Motivations and basic concepts Motivations
of multiple classifier systems. The “worst” case and “best” case motivations.
Practical and theoretical motivations. Basic concepts. Architectures for
multiple classifier systems. Ensemble types, combiner types. The concept of
classifier “diversity”. The design cycle of a multiple classifier system.
Creating multiple classifiers Systematic
methods for creating classifier ensembles. Methods
based on training data manipulation: data splitting methods, Bagging and
based on input and output feature manipulation: feature selection, the Random
Subspace method, noise injection, and error-correcting codes.
Combining multiple classifiers Methods
for combining multiple classifiers at the “abstract” level (voting methods, the
Behaviour Knowledge Space method, etc.) Methods
for combining multiple classifiers at the “rank” level (the Borda count method,
for combining multiple classifiers at the “measurement” level (linear
combiners, the product rule, etc.) Basic
concepts on dynamic classifier selection methods.