Browsing by Author "Tolun, MR"

Now showing 1 - 2 of 2

Citation - WoS: 1
Citation - Scopus: 1
A multi-relational rule discovery system
(Springer-verlag Berlin, 2003) Uludag, M; Tolun, MR; Etzold, T; 01. Atılım University
This paper describes a rule discovery system that has been developed as part of an ongoing research project. The system allows discovery of multi-relational rules using data from relational databases. The basic assumption of the system is that objects to be analyzed are stored in a set of tables. Multi-relational rules discovered would either be used in predicting an unknown object attribute value, or they can be used to see the hidden relationship between the objects' attribute values. The rule discovery system, developed, was designed to use data available from any possible 'connected' schema where tables concerned are connected by foreign keys. In order to have a reasonable performance, the 'hypotheses search' algorithm was implemented to allow construction of new hypotheses by refining previously constructed hypotheses, thereby avoiding the work of re-computing.
Citation - WoS: 2
Text Categorization With Ila
(Springer-verlag Berlin, 2003) Sever, H; Gorur, A; Tolun, MR; 01. Atılım University
The sudden expansion of the web and the use of the internet has caused some research fields to regain (or even increase) its old popularity. Of them, text categorization aims at developing a classification system for assigning a number of predefined topic codes to the documents based on the knowledge accumulated in the training process. We propose a framework based on an automatic inductive classifier, called ILA, for text categorization, though this attempt is not a novel approach to the information retrieval community. Our motivation are two folds. One is that there is still much to do for efficient and effective classifiers. The second is of ILA's (Inductive Learning Algorithm) well-known ability in capturing by canonical rules the distinctive features of text categories. Our results with respect to the Reuters 21578 corpus indicate (1) the reduction of features by information gain measurement down to 20 is essentially as good as the case where one would have more features; (2) recall/precision breakeven points of our algorithm without tuning over top 10 categories are comparable to other text categorization methods, namely similarity based matching, naive Bayes, Bayes nets, decision trees, linear support vector machines, steepest descent algorithm.