DYNAMIC FUZZY PATTERN RECOGNITION WITH APPLICATIONS TO FINANCE AND ENGINEERING LARISA ANGSTENBERGER Kluwer Academic Publishers Boston/Dordrecht/London
TABLE OF CONTENTS FOREWORD ACKNOWLEDGEMENTS XIX XXI 1 INTRODUCTION 1 1.1 GOALS AND TASKS OF THE BOOK 2 1.2 STRUCTURE OF THE BOOK 4 2 GENERAL FRAMEWORK OF DYNAMIC PATTERN RECOGNITION 7 2.1 THE KNOWLEDGE DISCOVERY PROCESS 7 2.2 THE PROBLEM OF PATTERN RECOGNITION 11 2.2.1 The process of pattern recognition 12 2.2.2 Classification of pattern recognition methods 14 2.2.3 Fuzzy pattern recognition 18 2.3 THE PROBLEM OF DYNAMIC PATTERN RECOGNITION 20 2.3.1 Mathematical description and modelling of dynamic systems 21 2.3.2 Terminology of dynamic pattern recognition 22 2.3.3 Goals and tasks of dynamic pattern recognition 32 3 STAGES OF THE DYNAMIC PATTERN RECOGNITION PROCESS 37 3.1 THE MONITORING PROCESS 37 3.1.1 Shewhart quality control charts 40 3.1.2 Fuzzy techniques for the monitoring process 43 3.1.2.1 Fuzzy quality control charts 44 3.1.2.2 Reject options in fuzzy pattern recognition 47 3.1.2.3 Parametric concept of a membership function for a dynamic classifier 50 3.2 THE ADAPTATION PROCESS 57 3.2.1 Re-learning of the classifier 58 3.2.2 Incremental updating of the classifier 60 3.2.3 Adaptation of the classifier 64 3.2.3.1 Learning from statistics approach 66 3.2.3.2 Learning with a moving time window 69 3.2.3.3 Learning with a template set 70 3.2.3.4 Learning with a record of usefulness 73 '3.2.3.5 Evaluation of approaches for the adaptation of a classifier 77
viii Table of Contents 4 DYNAMIC FUZZY CLASSIFIER DESIGN WITH POINT- PROTOTYPE BASED CLUSTERING ALGORITHMS 79 4.1 FORMULATION OF THE PROBLEM OF DYNAMIC CLUSTERING 80 4.2 REQUIREMENTS FOR A CLUSTERING ALGORITHM USED FOR DYNAMIC CLUSTERING AND CLASSIFICATION 86 4.3 DETECTION OF NEW CLUSTERS ' 88 4.3.1 Criteria for the detection of new clusters 89 4.3.2 Algorithm for the detection of new clusters 93 4.4 MERGING OF CLUSTERS 100 4.4.1 Criteria for merging of clusters 101 4.4.2 Criteria for merging of ellipsoidal clusters 106 4.4.3 Criteria and algorithm for merging spherical and ellipsoidal clusters 112 4.5 SPLITTING OF CLUSTERS 118 4.5.1 Criteria for splitting of clusters 118 4.5.2 Search for a characteristic pattern in the histogram 122 4.5.3 Algorithm for the detection of heterogeneous clusters to be split 125 4.6 DETECTION OF GRADUAL CHANGES IN THE CLUSTER STRUCTURE 132 4.7 ADAPTATION PROCEDURE 133 4.8 UPDATING THE TEMPLATE SET OF OBJECTS 136 4.8.1 Updating the template set after gradual changes in the cluster structure - 139 4.8.2 Updating the template set after abrupt changes in the cluster structure 142 4.9 CLUSTER VALIDITY MEASURES FOR DYNAMIC CLASSIFIERS 143 4.10 SUMMARY OF THE ALGORITHM FOR DYNAMIC FUZZY CLASSIFIER DESIGN AND CLASSIFICATION 151 5 SIMILARITY CONCEPTS FOR DYNAMIC OBJECTS IN PATTERN RECOGNITION 155 5.1 EXTRACTION OF CHARACTERISTIC VALUES FROM TRAJECTORIES 15 7 5.2 THE SIMILARITY NOTION FOR TRAJECTORIES & 160 5.2.1 Pointwise similarity measures 161 5.2.1.1 Choice of the membership function for the definition of pointwise similarity 165 5.2.1.2 Choice of the aggregation operator for the definition of pointwise similarity 169 5.2.2 Structural similarity measures 175 5.2.2.1 Similarity model using transformation functions 175 5.2.2.2 Similarity measures based on wavelet decomposition 177 5.2.2.3 Statistical measures of similarity 179
Table of Contents ix 5.2.2.4 Smoothing of trajectories before the analysis of their temporal behaviour 181 5.2.2.5 Similarity measures based on characteristics of trajectories 183 5.3 EXTENSION OF FUZZY PATTERN RECOGNITION METHODS BY APPLYING SIMILARITY MEASURES FOR TRAJECTORIES 197 6 APPLICATIONS OF DYNAMIC PATTERN RECOGNITION METHODS 199 6.1 BANKCUSTOMER SEGMENTATION BASEDON CUSTOMER BEHAVIOUR 199 6.1.1 Description of the credit data of bank customers 200 6.1.2 Goals of bank customer analysis 204 6.1.3 Parameter settings for dynamic classifier design and bank customer classification 206 6.1.4 Clustering of bank customers in Group ' Y' based on the whole temporal history of 24 months and using the pointwise similarity measure 208 6.1.5 Clustering of bank customers in Group 'N' based on the whole temporal history of 24 months and using the pointwise similarity measure 214 6.1.6 Segmentation of bank customers in Group ' Y' based on the partial temporal history and using the pointwise similarity measure 220 6.1.7 Clustering of bank customers in Group 'N' based on partial temporal history and using the pointwise similarity measure 226 6.1.8 Comparison of clustering results for customers in Groups ' Y' and 'N' 232 6.2 COMPUTER NETWORK OPTIMISATION BASED ON DYNAMIC NETWORK LOAD CLASSIFICATION 233 6.2.1 Data transmission in computer networks 234 6.2.2 Data acquisition and pre-processing for the network analysis 238 6.2.3 Goals of the analysis of load in a computer network 242 6.2.4 Parameter settings for dynamic classifier design and classification of network traffic 243 6.2.5 Recognition of typical load states in a qpmputer network using the pointwise similarity measure 244 6.2.6 Recognition of typical load states in computer network using the structural similarity measure 254 7 CONCLUSIONS 265 REFERENCES 269
Table of Contents APPENDIX 279 UNSUPERVISED OPTIMAL FUZZY CLUSTERING ALGORITHM OF GATH AND GEVA 279 DESCRIPTION OF IMPLEMENTED SOFTWARE 282 INDEX, 285