Accepted Papers
- OUDG: Cross model datum access with semantic preservation for legacy databases
Joseph Fong and Kenneth Wong, City University of Hong Kong, Hong Kong.
ABSTRACT
Conventional databases are associated with a plurality of database models. Generally database models
are distinct and not interoperable. Data stored in a database under a particular database model can be
termed as “siloed data”. Accordingly, a DBMS associated with a database silo, is generally not interoperable with another database management system associated with another database sil. This can
limit the exchange of information stored in a database where those desiring to access the information are
not employing a database management system associated with the database model related to the information. The DBMS of various data models have proliferated into many companies, and become their
legacy databases. There is a need to access these legacy databases using ODBC. An ODBC is for the users to transform a legacy database into another legacy database. This paper offers an end user’s tool of Open Universal Database Gateway(OUDG) to replace ODBC by transforming a source legacy database data into Flattened XML documents, and further transform Flattened XML document into a target legacy database. The Flattened XML document is a mixture of relational and XML data models, which is user friendly and is a data standard on the Internet. The result of reengineering legacy databases into each other through OUDG is information lossless by the preservation of their data semantics in terms of data dependencies.
- Class Discovery by Semi-supervised Multi-label Classi_cation
Yuichiro Kase and Takao Miura, HOSEI University, Japan.
ABSTRACT
In this investigation, we propose a novel approach to find potential classes in news documents. Our basic idea comes from a fact there exist close relationship between new classes and probability vectors of multiple labeling of the documents. By classifying document articles with multiple labels, we obtain probability distribution function to each label by means of semi-supervised learning: with small amount of articles considered as training data, we extract probability vectors of several labels. Here we assume multinomial distribution over words, and apply EM algorithm to obtain membership distribution over linear mixture models. Then we apply clustering approach in the label space to find potential classes. We discuss some experimental results to show how well the proposed approach works.