Similarity-Based Clustering

Similarity-based learning methods have a great potential as an intuitive and ?exible toolbox for mining, visualization,and inspection of largedata sets. They combine simple and human-understandable principles, such as distance-based classi?cation, prototypes, or Hebbian learning, with a large variety of di?erent, problem-adapted design choices, such as a data-optimum topology, similarity measure, or learning mode. In medicine, biology, and medical bioinformatics, more and more data arise from clinical measurements such as EEG or fMRI studies for monitoring brain activity, mass spectrometry data for the detection of proteins, peptides and composites, or microarray pro?les for the analysis of gene expressions. Typically, data are high-dimensional, noisy, and very hard to inspect using classic (e. g. , symbolic or linear) methods. At the same time, new technologies ranging from the possibility of a very high resolution of spectra to high-throughput screening for microarray data are rapidly developing and carry thepromiseofane?cient,cheap,andautomaticgatheringoftonsofhigh-quality data with large information potential. Thus, there is a need for appropriate - chine learning methods which help to automatically extract and interpret the relevant parts of this information and which, eventually, help to enable und- standingofbiologicalsystems,reliablediagnosisoffaults,andtherapyofdiseases such as cancer based on this information. Moreover, these application scenarios pose fundamental and qualitatively new challenges to the learning systems - cause of the speci?cs of the data and learning tasks. Since these characteristics are particularly pronounced within the medical domain, but not limited to it and of principled interest, this research topic opens the way towardimportant new directions of algorithmic design and accompanying theory.

Autorentext
Die vorliegende Arbeit entstand im wesentlichen am Institut für Informatik der Universität Leipzig in der Forschungsgruppe Neuroinformatik, der der Autor von 1991 bis 1994 angehörte. Sie wurde 1996 als Dissertationsschrift angenommen, der Betreuer war Dr. habil. Ralf Der. Neben der Theorie und Anwendung neuronaler Netze arbeitet der Autor auf den Gebieten Genetische Algorithmen, maschinelle Lernverfahren, Optimierung und Statistik.

Klappentext

This book is the outcome of the Dagstuhl Seminar on "Similarity-Based Clustering" held at Dagstuhl Castle, Germany, in Spring 2007.

In three chapters, the three fundamental aspects of a theoretical background, the representation of data and their connection to algorithms, and particular challenging applications are considered. Topics discussed concern a theoretical investigation and foundation of prototype based learning algorithms, the development and extension of models to directions such as general data structures and the application for the domain of medicine and biology.

Similarity based methods find widespread applications in diverse application domains, including biomedical problems, but also in remote sensing, geoscience or other technical domains. The presentations give a good overview about important research results in similarity-based learning, whereby the character of overview articles with references to correlated research articles makes the contributions particularly suited for a first reading concerning these topics.

Inhalt
I: Dynamics of Similarity-Based Clustering.- Statistical Mechanics of On-line Learning.- Some Theoretical Aspects of the Neural Gas Vector Quantizer.- Immediate Reward Reinforcement Learning for Clustering and Topology Preserving Mappings.- II: Information Representation.- Advances in Feature Selection with Mutual Information.- Unleashing Pearson Correlation for Faithful Analysis of Biomedical Data.- Median Topographic Maps for Biomedical Data Sets.- Visualization of Structured Data via Generative Probabilistic Modeling.- III: Particular Challenges in Applications.- Learning Highly Structured Manifolds: Harnessing the Power of SOMs.- Estimation of Boar Sperm Status Using Intracellular Density Distribution in Grey Level Images.- HIV-1 Drug Resistance Prediction and Therapy Optimization: A Case Study for the Application of Classification and Clustering Methods.