S+SSPR, a satellite event of the ICPR, took place last week in the hustle and bustle of the University of Central Florida in Orlando.
Dr. Tin Kam Ho did a talk at S+SSPR entitled Data complexity analysis: Linkage between context and solution in classification. She explained that classification accuracy depends on both learner quality and intrinsic data complexity. So, changes in source data and feature transformations are necessary to reduce the data difficulty and to improve learner performance. Among her collaborators, she mentioned the work performed by Ester Bernadó, Albert Orriols, and Núria Macià. Data complexity is a large knowledge area which should be explored in many directions, Dr. Ho suggested some of them for further research, easier to follow than the conference indications.
This picture was taken by Dr. Ludmila Kuncheva, another invited speaker. With some nice pictures and a fresh touch of humor, she presented Linear discriminant classifier (LDC) for streaming data with concept drift. She provided a framework within which theoretical relationship can be sought between the window size and the classification error.
She is also offering a promising lecture in the ICPR under the interesting title Classifier ensembles: Facts, fiction, faults and future.
Prof. Pedro Domingos presented an impressive work summarized by the title Markov logic: A unifying language for structural and statistical pattern recognition. It is a powerful language that combines statistical and structural aspects of the input data. Models in Markov logic are sets of weighted formulas in first-order logic which are interpreted as templates for features of Markov random fields.
Finally, Dr. Horst Bunke presented Graph classification on dissimilarity space embedding. He proposed a methodology that consists in embedding graphs in vector spaces and then applying a statistical classifier to outperform classifiers that directly operate in the graph domain. It is a contribution towards unifying the domains of structural and statistical pattern recognition. We highlight a graph data base repository that contains data sets covering a wide spectrum of different applications.
For further information, please refer to S+SSPR web site.