Abstract
dc:descriptionThe fast growing of the Internet and the continuous rising of the available material in digital form, make ergent the need for classification algorithms. Older techniques of the Knowledge Engineering back at the '80s, have been replaced by Machine Learning techniques. The last 15 years a large number of methods have been studied and developed, which have succeed in solving this problem. Such techniques include Decision Tress, Naive Bayes, Neural Networks, Linear Classifiers, Logistic Regression, Perceptron, Support Vector Machines etc. An important group of classifiers is Linear Classifiers, which try to classify data by defining separating hyperplanes. Types of Linear Classifiers are Centroid Classifier, Rocchio Classifier and Perceptron Classifier. By combining characteristics of those three simple classifiers, a new, fast and accurate classifier is defined, under the symbolic name Modified Perceptron, due to its resemblence with the classic Perceptron Classifier. This new classifier is proved that it converges and it is shown experimentally that it converges much faster than other Linear Classifiers. By evaluating its performance on classifying international text corpuses and corpuses from conference challenges, it turned that its performance is comparable, if not better, than the state of the art in the field of Text Classification, such as SVMs. Note that it took the first place in the international challenge on spam detection on social bookmarking networks of ECML 2008. Evaluation is performed by the classic model of flat classification, where each category is considered independent from others. This technique, one versus all, has some limitations, such as how it scales when the number of categories is very large or the number of examples is very large. This limitation refers in space requirements, beacause all classifiers have to be stored in memory. As for time requirements, time complexity is O(NM), where N is the number classifiers and M is the dimensionality of each example. To overcome those limitation, a hierarchical model for classification has been developed. Relations among categories where defined and an experimental search was done on the sampling for creating negative sets for each category. The proposed algorithm was succesfully applied on very large classification problems and new issues for further research and improvements where opened.
Degree
thesis:*- Grantor dc:publisher
- Athens University Economics and Business (AUEB)
- Year dc:date
- 2011
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Gkanogiannis, Anestis
- Γκανόγιαννης, Ανέστης
Subjects
dc:subject × 13Rights
- Language dc:language
- gre
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/24406