Multiple Kernel Clustering
Overview
Clustering is an important unsupervised learning method for dividing data into a set of disjoint groups with high intra-cluster similarity and low inter-cluster similarity.
Most clustering algorithms assume a single affinity matrix recording pairwise similarity between data.
However, in many applications, there could be multiple potentially useful features and thereby multiple affinity matrices.
For better clustering results, multiple affinity matrices should be aggregated or fused.
However, careless aggregation might make even worse clustering results. We applied the multiple kernel learning theory to fuzzy clustering (published as an IEEE TFS paper) and spectral clustering (ICASSP 2012 and CVPR 2012) so that they simultaneously seek for an optimal combination of affinity matrices and optimize clustering results.
The resultant algorithms are more immune to ineffective affinities and irrelevant features. We have applied these algorithms to concept-based image clustering, face image clustering and text clustering.
Publications
- Affinity Aggregation
for Spectral Clustering
-
Hsin-Chien Huang,
Yung-Yu Chuang,
Chu-Song Chen
- CVPR 2012
- Multi-Affinity
Spectral Clustering
-
Hsin-Chien Huang,
Yung-Yu Chuang,
Chu-Song Chen
- ICASSP 2012
- Multiple
Kernel Fuzzy Clustering
-
Hsin-Chien Huang,
Yung-Yu Chuang,
Chu-Song Chen
- IEEE TFS 2012
Support
This research is supported by:
- NSC100-2628-E-002-009
- NSC100-2622-E-002-016-CC2
- NSC98-2221-E-001-012-MY3
cyy -a-t- csie.ntu.edu.tw