TY - GEN
T1 - A generalized maximum entropy approach to Bregman co-clustering and matrix approximation
AU - Banerjee, Arindam
AU - Dhillon, Inderjit
AU - Ghosh, Joydeep
AU - Merugu, Srujana
AU - Modha, Dharmendra S.
PY - 2004
Y1 - 2004
N2 - Co-clustering is a powerful data mining technique with varied applications such as text clustering, microarray analysis and recommender systems. Recently, an information-theoretic co-clustering approach applicable to empirical joint probability distributions was proposed. In many situations, co-clustering of more general matrices is desired. In this paper, we present a substantially generalized co-clustering framework wherein any Bregman divergence can be used in the objective function, and various conditional expectation based constraints can be considered based on the statistics that need to be preserved. Analysis of the co-clustering problem leads to the minimum Bregman information principle, which generalizes the maximum entropy principle, and yields an elegant meta algorithm that is guaranteed to achieve local optimality. Our methodology yields new algorithms and also encompasses several previously known clustering and co-clustering algorithms based on alternate minimization.
AB - Co-clustering is a powerful data mining technique with varied applications such as text clustering, microarray analysis and recommender systems. Recently, an information-theoretic co-clustering approach applicable to empirical joint probability distributions was proposed. In many situations, co-clustering of more general matrices is desired. In this paper, we present a substantially generalized co-clustering framework wherein any Bregman divergence can be used in the objective function, and various conditional expectation based constraints can be considered based on the statistics that need to be preserved. Analysis of the co-clustering problem leads to the minimum Bregman information principle, which generalizes the maximum entropy principle, and yields an elegant meta algorithm that is guaranteed to achieve local optimality. Our methodology yields new algorithms and also encompasses several previously known clustering and co-clustering algorithms based on alternate minimization.
KW - Bregman divergences
KW - Co-clustering
KW - Matrix Approximation
UR - http://www.scopus.com/inward/record.url?scp=12244261221&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=12244261221&partnerID=8YFLogxK
U2 - 10.1145/1014052.1014111
DO - 10.1145/1014052.1014111
M3 - Conference contribution
AN - SCOPUS:12244261221
SN - 1581138881
SN - 9781581138887
T3 - KDD-2004 - Proceedings of the Tenth ACM SIGKDD International Conference on Knowledge Discovery and Data Mining
SP - 509
EP - 514
BT - KDD-2004 - Proceedings of the Tenth ACM SIGKDD International Conference on Knowledge Discovery and Data Mining
PB - Association for Computing Machinery (ACM)
T2 - KDD-2004 - Proceedings of the Tenth ACM SIGKDD International Conference on Knowledge Discovery and Data Mining
Y2 - 22 August 2004 through 25 August 2004
ER -