Campus Event Calendar
Event Entry
What and Who
Link Propagation: A Semi-supervised Learning Algorithm for Link Prediction
Dr. Hisashi Kashima
IBM Research Japan
Talk
Hisashi Kashima has been a research staff member of Data Analytics Group in Tokyo
Research Laboratory of IBM Research since April, 1999.
He is working on machine learning and data mining research and their applications to
bioinformatics, autonomic computing, and business and industrial analytics.
His previous research works include development of kernel methods for structured data
such as trees and graphs, predictive modeling of biological/social networks, and
anomaly detection for industrial systems, and those contributions were awarded by
academic societies.
He also takes advantage of his machine learning skills in businesses, and has many
issued/disclosed patents.
He obtained his B.S. degree in applied mathematics and physics in 1997, and a M.S.
degree in systems engineering in 1999, and a Ph.D. degree in informatics in 2007 from
Kyoto University in Japan.
AG 1, AG 3, AG 5, RG2, AG 2, AG 4, RG1, SWS
Public Audience
English
Note: We use this to send email in the morning.
Date, Time and Location
Wednesday, 17 September 2008
11:00
45 Minutes
E1 4
023
Saarbrücken
Abstract
We present a new semi-supervised learning algorithm for link prediction
problems, where the task is to
predict unknown parts of the network structure by using auxiliary
information such as node
similarities.
Although the link prediction problem can be seen as the problem of
completing an adjacency matrix, we
add another dimension to the adjacency matrix, which results in a
third-order tensor completion
problem.
The tensor representation of the network structure allows us to handle not
only the existence of links
but also the types of links, including temporal links.
Also, the tensor representation makes it possible to simultaneously
predict multiple networks that
have correlations with each other.
We use one of the well-known approaches of semi-supervised learning called
label propagation for link
prediction.
Since naive application of label propagation to link prediction causes
scalability problems, we
propose using the Kronecker product similarity and the Kronecker sum
similarity as the similarity
matrices used in label propagation.
We propose an efficient algorithm based on the conjugate gradient method
exploiting the structure
ofthe similarity matrices.
problems, where the task is to
predict unknown parts of the network structure by using auxiliary
information such as node
similarities.
Although the link prediction problem can be seen as the problem of
completing an adjacency matrix, we
add another dimension to the adjacency matrix, which results in a
third-order tensor completion
problem.
The tensor representation of the network structure allows us to handle not
only the existence of links
but also the types of links, including temporal links.
Also, the tensor representation makes it possible to simultaneously
predict multiple networks that
have correlations with each other.
We use one of the well-known approaches of semi-supervised learning called
label propagation for link
prediction.
Since naive application of label propagation to link prediction causes
scalability problems, we
propose using the Kronecker product similarity and the Kronecker sum
similarity as the similarity
matrices used in label propagation.
We propose an efficient algorithm based on the conjugate gradient method
exploiting the structure
ofthe similarity matrices.
Contact
Hiroto Saigo
+49 681 9325 319
--email hidden
passcode not visible
logged in users only
Hiroto Saigo, 09/13/2008 22:35 -- Created document.