Aristotle University Of Thessaloniki (AUTH)
Εξόρυξη γνώσης από βάσεις χρονοσειρών: επιλογή χαρακτηριστικών και κατηγοριοποίηση
Abstract
dc:descriptionThe objective of the thesis is to develop and apply data mining techniques in time series databases, and in particular to investigate for the most relevant time series measures to be used for feature based clustering and classification. In the course of this work, the following goals were achieved: 1) a large number of features used in time series analysis were computed, 2) new algorithms have been developed for optimal feature subset selection, 3) they have been compared to other currently favored feature selection methods on simulated and real world databases, especially epileptic electroencephalograms (EEG), 4) the proposed methods were shown to be able to find a small subset of the most useful features for classification purposes, and particularly for epileptic EEG it was shown that such a feature subset can possibly play the role of "biomarker" and distinguish the late preictal stage, and finally 5) two menu-based Matlab modules were developed, the first for the computation of a large number of measures on time series databases, and the second for feature selection, and feature-based classification and clustering. In this thesis, data mining techniques were applied on time series using a variety of measures (linear, nonlinear, oscillation related measures, and other measures) that were calculated on each time series and are supposed each to capture a different characteristic of the time series. In the study we considered various sets of features (10-312 features), including the most known measures of the analysis of scalar time series, especially the measures used in EEG analysis. Our results showed that oscillation related features are among the first to be included in the optimal feature subset. The first problem we investigated in the search of best features was the effect of standardization and normalization methods on the feature selection and subsequently the performance of the classification. We evaluated different standardization methods and proposed a new method, the transform to Gaussian distribution, which turned out to have good scaling and stability properties, and could better handle outliers. We developed two feature selection schemes that derive small and optimal feature subsets. The first scheme is called Forward Sequential Selection with CRI (FSSwCRI) and is wrapper based on the k-means clustering algorithm in order to derive the feature subset that classifies best the time series database, according to the criterion of the corrected Rand index. The second feature selection scheme we developed is a conditional mutual information filter based on nearest neighbour estimate and is called CMINN. Applications of CMINN to simulations and to ten benchmark databases showed that CMINN is superior compared to mRMR and MaxiMin and derive feature subsets of smaller cardinality. FSSwCRI was applied to classification of preictal states in epileptic EEG databases, were it was compared to Support Vector Machines with Recursive Feature Elimination (SVM-RFE) and to a simple Mutual Information (MI) filter. We concluded that a global feature subset of the 16 most frequent features selected by FFSSwCRI can play the role of a biomarker and distinguish early and late preictal states.
Degree
thesis:*- Grantor dc:publisher
- Aristotle University Of Thessaloniki (AUTH)
- Year dc:date
- 2012
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Τσιμπίρης, Αλκιβιάδης
- Tsimpiris, Alkiviadis
Subjects
dc:subject × 20- Εξόρυξη γνώσης
- Κατηγοριοποίηση
- Χρονοσειρές
- Επιλογή Χαρακτηριστικών
- Συσταδοποίηση
- Αμοιβαία πληροφορία
- Φίλτρο και περιτύλιγμα
- Ηλεκτροεγκεφαλογράφημα
- Data mining
- Classification
- Time series
- Feature Selection
- Clustering
- Mutual information
- Filter and wrapper
- Electroencephalography EEG
- Natural Sciences
- Computer and Information Sciences
- Φυσικές Επιστήμες
- Επιστήμη Ηλεκτρονικών Υπολογιστών και Πληροφορική
Rights
- Language dc:language
- gre
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/27116