TY - GEN
T1 - Robust integration of multiple information sources by view completion
AU - Subramanya, Shankara
AU - Li, Baoxin
AU - Liu, Huan
PY - 2008
Y1 - 2008
N2 - There are many applications where multiple data sources, each with its own features, are integrated in order to perform an inference task in an optimal way. Researchers have shown that for many tasks like webpage classification, image classification, and pattern recognition, combining data from multiple information sources yields significantly better results than using a single source. In these tasks each of the multiple data sources can be thought of as providing one view of the underlying object. However in many domains not all of the views are available for the available instances; some of the views would be missing. This problem of missing views affects the performance of the machine learning task. In this paper we provide a method of view completion to heuristically predict the missing views. We show that with view completion we are able to achieve significantly better results. We also show that by considering the information at a higher level in terms of views rather than considering them at a lower level in terms of features we are able to achieve better results. We demonstrate this by comparing our method with existing methods which consider the missing views problem as a missing value problem.
AB - There are many applications where multiple data sources, each with its own features, are integrated in order to perform an inference task in an optimal way. Researchers have shown that for many tasks like webpage classification, image classification, and pattern recognition, combining data from multiple information sources yields significantly better results than using a single source. In these tasks each of the multiple data sources can be thought of as providing one view of the underlying object. However in many domains not all of the views are available for the available instances; some of the views would be missing. This problem of missing views affects the performance of the machine learning task. In this paper we provide a method of view completion to heuristically predict the missing views. We show that with view completion we are able to achieve significantly better results. We also show that by considering the information at a higher level in terms of views rather than considering them at a lower level in terms of features we are able to achieve better results. We demonstrate this by comparing our method with existing methods which consider the missing views problem as a missing value problem.
UR - http://www.scopus.com/inward/record.url?scp=51949091936&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=51949091936&partnerID=8YFLogxK
U2 - 10.1109/IRI.2008.4583064
DO - 10.1109/IRI.2008.4583064
M3 - Conference contribution
AN - SCOPUS:51949091936
SN - 9781424426607
T3 - 2008 IEEE International Conference on Information Reuse and Integration, IEEE IRI-2008
SP - 398
EP - 403
BT - 2008 IEEE International Conference on Information Reuse and Integration, IEEE IRI-2008
T2 - 2008 IEEE International Conference on Information Reuse and Integration, IEEE IRI-2008
Y2 - 13 July 2008 through 15 July 2008
ER -