Transactions on Machine Learning and Data Mining

Special issue of “Transactions on Machine Learning and Data Mining” presents three
carefully selected papers from World Congress “The Frontiers in Intelligent Data and
Signal Analysis” DSA 2011 - one paper form MLDM 2011and two papers from
ICDM 2011. The papers make significant contributions to the fields of data mining
and machine learning.
The first paper presents [1] a new algorithms for discovering changes in the
correlations of an itemset among contrasted databases to detect potential changes
among them. This is accomplished by a simple double-clique algorithm that
enumerates itemsets that show higher, but not too high, correlation in one database
and lower correlation in another one for contrasting. The itemset correlation is
measured by k-way mutual information. The authors are looking for cases when weak
correlation is present in the first database and moderate correlation in the second
database, while there is a minimum correlation increase from the first to the second
database. The main contribution of the paper lies in proposing an efficient algorithm
for identifying itemsets satisfying correlation constraints. This is accomplished by
enumerating double-cliques in anti-correlation graphs. In simulation experiments the
proposed algorithm performs almost two orders of magnitudes faster than the naive
approach.
The second paper [2] presents a novel technique for agricultural monitoring by
mining Satellite Image Time Series over agriculture cultivated areas. It uses Grouped
Frequent Sequential patterns (GFS-patterns) extended to spatio-temporal context in
order to extract sets of connected pixels sharing a similar temporal evolution. It
allows to uncover sets of pixels satisfying two properties of cultivated areas, namely
being spatially connected and sharing similar temporal evolutions. No prior
knowledge of the identified regions is assumed and no distance measures are needed.
The general framework of GFS-patterns is extended in two directions. Firstly, the
connectivity constraint is used in the search space exploration leading to significant
reduction of execution times on real Satellite Image Time Series of cultivated areas.
Secondly, simple post-processing with a maximality constraint over the patterns
significantly improves efficiency. The experiments were performed on database of
real images ADAM (Data Assimilation by Agro-Modeling) of SITS (Centre National
d’Etudes Spatiales (2010)) demonstrate that pushing the average connectivity measure
constraint, during GFS-pattern extraction is effective to reduce the search space. It is
also demonstrated that together with a maximality constraint, the proposed approach
is useful to find meaningful patterns in real data.
The third paper [3] proposes a classification system for recognition of wood
species based on microscopic images of wood pores. The system introduces two new
sets of features invariant to rotations, scale and translations. The features are based on
the nearest pore pairs and on pore diameter change distributions. The machine
learning algorithm C4.5 was used to generate decision trees and decision rules. In
simulation experiments with real wood data 83.7% classification rates were obtained
which compare favorably with the performance of human experts.
All three papers provide novel and significant contributions to the field of data
mining and machine learning. All proposed algorithms are tested on real data and they
perform well. The methodology and algorithms presented in the papers may be
applied in other application domains and thus have lasting value. The editors ought to
be commanded for choosing valuable and important contributions for the special
issue.