Learning with Multiple Graphs


Friedrich-Miescher-Laboratory, Max-Planck-Society, Germany

2005. 3. 30


In some application area, it is more natural to represent domain knowledge using graphs. And frequently there exist multiple descriptions of graphs for the same set of nodes depending on various kinds of information sources, but no single graph is sufficient to reliably predict class labels of unlabelled nodes.

One way to enhance the reliability is integrating multiple graphs since individual graphs are partly independent and partly complementary to each other for prediction.

The talk will present two methods for integrating multiple graphs within a framework of semi-supervised learning.

(A) One method alternates between minimizing the objective function with respect to network output and with respect to combining weights.

(B) The other method finds the combining weights by convex optimization (no local minima).

The methods have been applied to the task of protein functional class prediction in yeast. Both methods perform significantly better than the same algorithm trained on any single graph. Particularly, the method (B) shows a comparable accuracy to the SDP/SVM in a remarkably short time.


This page is maintained by Ji-seon Yoo (jsyoo@bi.snu.ac.kr).
Last update: March 24, 2005