Computational and Statistical Learning Group CASTLE

1
Computational and Statistical Learning Group
(CASTLE)

• Professor Edgar Acuna
• Mathematics Department
• University of Puerto Rico at Mayaguez
• Partially supported by ONR Grant N0014-03-0359
  

2
Our research is focused in the development of
computational and statistical methods for
knowledge discovery in databases related to
Pattern Recognition, Machine Learning,
Bioinformatics, and Data Mining
3
CASTLEs Members1
4
CASTLEs Members2
5
CASTLEs Alummi

• Alex Rojas (MS 2001, now at CMU)
• Adriana Lopez (MS 2002-now at U. Pittsburgh)
• Jose Vega (PhD 2004, now at the UPR-Medical
  School)
• Santiago Velasco (MS 2004, now at UPRM

6
CASTLEs collaborators

• Ana Patricia Ortiz, Puerto Rico Cancer Center.
• Idhaliz Flores, Ponce Medical School, Puerto
  Rico.
• Jose Vega ( University of Puerto Rico-School of
  Medicine)

7
Research topics

• Data preprocessing
• Treatment of missing values in Data Mining
• Normalization in Data Mining
• Feature selection procedures for high dimensional
  data Wrappers, Filters and hybrid methods with
  application to Bionformatics. Feature selection
  based on rough sets.
• Feature extraction procedures for high
  dimensional data, such as Partial least Squares
  and supervised Principal components, with
  application to Bioniformatics and Chemometrics.
• Instance selection procedures, such as
  progressive sampling, to make possible the
  application of statistical methods and machine
  learning techniques.

8
Research topics 2

• Regularization Methods. Use of shrunken
  estimators for logistic regression trying to
  work with classifiers that use all the features
  without performing feature selection and feature
  extraction. Application in Bioinformatics and
  Chemometrics.
• Outlier detection procedures, such as
  distance-based outliers and density local
  outliers. We will apply these procedures to
  network intrusion detection.
• Visualization procedures for Data Mining. We are
  enhancing graphics such as parallel coordinate
  plots, surveyplot, and star coordinates that will
  allow us to perform exploratory data analysis
  previously to the application of a knowledge
  discovery technique.

9
Research Topics3

• Parallel Data Mining. The use of parallel
  computation will allow us to deal with very large
  datasets. In particular we are developing
  parallel algorithms to perform data preprocessing
  tasks in very large datasets. Visualization and
  computation of meta-classifiers using parallelism
  is also being considered.
• Unsupervised learning to find features that
  behave similarly in various conditions and to
  find subgroups of instances that similar to each
  other. In particular we are interested in
  validation of clustering algorithms.
• We are investigating extensions of data mining
  tasks to the multi-relational case (multiple
  tables).
• Interface between the R statistical computing
  language and SQL in order to manipulate large
  datasets.
• Building of a visual event programming on R to
  perform mainly data preprocessing tasks.

10
Lab Equipment

• 155 square feet room.
• A cluster of five dual Pentium Xeon processors
  Dell workstations. Each of them running at 3.06
  GHz and with 3MB of RAM memory.
• A Color laserjet 4050 HP printer

11
Recent Publications1

• 1 Lozano, E. and Acuna, E. (2005). A 3D
  extension of the Star coordinates display.
• 2 Lozano, E. and Acuna, E. (2005) Parallel
  algoritms for distance-based and density
  local-based outliers. To appear in the
  Proceedings of the ICDM05 organized by the IEEE.
  
• 3 Vega, J. and Acuna, E. (2005).
  Generalizations of PLS for dimensionality
  reduction in supervised classification.
  Proceedings of the Fourth International
  Conference in Statistics and Related fields.
  Hawaii.
• 4 Acuna, E. and Rodriguez, C. (2004). The
  effect of outliers on the misclassification error
  rate. Submitted to the IEEE Transactions on
  Knowledge and Data Engineering.
• 5 Acuna, E. and Rodriguez, C. (2004). The
  treatment of missing values and its effect in the
  classifier accuracy. In D. Banks, L. House, F.R.
  McMorris, P. Arabie, W. Gaul (Eds).Classification,
  Clustering and Data Mining Applications.
  Springer-Verlag Berlin-Heidelberg, 639-648.

12
Recent Publications2

• 6 Acuña E. and Coaquira, F. (2003). A
  comparison of feature selection procedures for
  classifiers based on kernel density estimation.
  Proceedings of the International conference on
  computer, communication and control technologies,
  CCCT03. Vol I. p. 462-467. Orlando, Florida.
• 7 Daza, L. and Acuña, E.. (2003) Combining
  classifiers based on Gaussian Mixtures.
  Proceedings of the International conference on
  computer, communication and control technologies,
  CCCT03. Vol I. p. 473-478. Orlando, Florida.
• 8 Lozano, E. and Acuña, E. (2002) Parallel
  computation of kernel density estimates
  classifiers and their ensembles. Proceedings of
  the International conference on computer,
  communication and control technologies, CCCT03.
  Vol I. p. 473-478. Orlando, Florida.

13
Recent Publications3

• 9 Acuña, E , (2003) A comparison of filters and
  wrappers for feature selection in supervised
  classification. Proceedings of the Interface
  2003 Computing Science and Statistics. Vol 34.
• 10 Acuña, E. Rojas, A., and Coaquira, F.
  (2002). The effect of feature selection on
  combining classifiers based on kernel density
  estimates. In K. Jajuga, A. Sokodowski, H.-H Bock
  (Eds). Classification, Clustering and Data
  Analysis. Springer, Heidelberg, 161-168.
• 11 Acuña, E., (2002) Combining Classifiers
  based on Kernel density classifiers and Gaussian
  mixtures. Proceedings of the Interface 2002
  Computing Science and Statistics. Vol 33.

14
Software

• The Dprep package A library of 68 R functions to
  perform mainly data preprocessing tasks
  including range normalization, discretization,
  handling of missing values, outlier detection,
  feature selection, and visualization.

15
Castle Members
Elio Lozano Parallel Data mining
Caroline Rodriguez (PR) Data preprocessing-Visual
ization
16
Castle Members
Frida Coaquira Rough sets for KDD
Luis Daza Instance Selection
17
Castle Members
Trilce Encarnation Databases
Marggie Gonzalez Cluster Validation
18
Castle Members

Jaime Porras Supervised Principal component for
classification
Karen Prieto Shrunken estimators for logistic
regression
19
Castle Members
Carlos Lopez Bayesian networks classifiers
Carmen Saldana Statistics
20
Castle Members
Roxana Aparicio Databases
Sindy Diaz Statistics