Μεθοδολογίες μηχανικής μάθησης βασισμένες σε τεχνητά ανοσοποιητικά συστήματα
Artificial immune system based machine learning methodologies
Σωτηρόπουλος, Διονύσιος Ν.
KeywordsΑναγνώριση προτύπων ; Τεχνητά ανοσοποιητικά συστήματα ; Γενετικοί αλγόριθμοι ; Μηχανική μάθηση ; Ταξινόμηση ; Pattern recognition ; Artificial Immune Systems ; Genetic algorithms ; Machine learning ; Classification
The current Ph.D thesis lies within the field of Pattern Recognition, providing theoretical and experimental justifications concerning the validity of Artificial Immune Systems as an alternative machine learning paradigm. The main source of inspiration stems from the fact that the Adaptive Immune System constitutes one of the most sophisticated biological systems that is particularly evolved in order to continuously address an extremely unbalanced pattern classification problem, by performing the self / non-self discrimination process. The primary effort undertaken in this dissertation is focused on addressing the fundamental problems of Pattern Recognition by developing Artificial Immune System-based machine learning algorithms. Therefore, the relevant research is particularly interested in providing alternative machine learning approaches for the problems of Clustering, Classification and One-Class Classification, measuring their efficiency against state of the art pattern recognition paradigms such as the Support Vector Machines. Pattern classification is specifically studied within the context of the Class Imbalance Problem dealing with extremely skewed training data sets. Specifically, the experimental results presented in this thesis involve degenerated binary classification problems where the class of interest to be recognized is known through a limited number of positive training instances. In other words, the target class occupies only a negligible volume of the entire pattern space while the complementary space of negative patterns remains completely unknown during the training process. Therefore, the effect of the Class Imbalance Problem on the performance of the proposed Artificial Immune System-based classification algorithm constitutes one of the secondary objectives of this thesis. The general experimentation framework adopted throughout the current dissertation in order to assess the efficiency of the proposed clustering, classification and one class classification algorithms was an open collection of one thousand (1000) pieces from 10 classes of western music. This collection, in particular, has been extensively used in applications concerning music information retrieval and music genre classification [221, 135]. The following list summarizes the pattern recognition problems addressed in the current Ph.D thesis through the application of specifically designed Artificial Immune System-based machine learning algorithms: 1. Artificial Immune System-Based music piece clustering and Database Organization [202, 201]. 2. Artificial Immune System-Based Customer Data Clustering in an e-Shopping application . 3. Artificial Immune System-Based Music Genre Classification . 4. A Music Recommender Based on Artificial Immune Systems . Specifically, the experimental results presented in this thesis demonstrate that the general framework of Artificial Immune Systems constitutes a valid machine learning paradigm.