180 likes | 216 Views
This study introduces a context-aware clustering method for discovering higher-level compositional patterns through co-occurrences. The proposed approach efficiently incorporates spatial contexts to enhance primitive feature clustering. By utilizing a nested-EM solution, the algorithm ensures convergence in finite steps, with applications in image pattern discovery and multiple-feature clustering. It can be adapted to various clustering challenges.
E N D
Context-Aware Clustering Junsong Yuan and Ying Wu EECS Dept., Northwestern University
Contextual pattern and co-occurrences ? Spatial contexts provide useful cues for clustering
K-means revisit EM Update Binary label indicator Assumption: data samples are independent Limitation: contextual information of spatial dependency is not considered in clustering data samples
Clustering higher-level patterns • Regularized k-means Not a smooth term! Distortion in hamming space charactering contextual patterns Distortion in original feature space Same as traditional K-means clustering Regularization term due to contextual patterns
Chicken and Egg Problem • Hamming distance in clustering contextual patterns • Matrix form J1 is coupled with J2 • Cannot minimize J1 and J2 separately !
Decoupling Fix Update Fix Update
Nested-EM solution Nested E-step M-step Update and separately Theorem of convergence the nested-EM algorithm can converge in finite steps.
K-means Initialization
1st round Final Phrases
Multiple-feature clustering • Dataset: handwritten numerical (‘0’-‘9’) from UCI data set • Each digit has three different types of features • Contextual pattern corresponds to compositional feature • Different types of features serve as contexts of each other • Clustering each type of features into 10 “words” • Clustering 10 “phrases” based on a word-lexicon of size 3x10
Conclusion • A context-aware clustering formulation proposed • Targets on higher-level compositional patterns in terms of co-occurrences • Discovered contextual patterns can feed back to improve the primitive feature clustering • An efficient nested-EM solution which is guaranteed to converge in finite steps • Successful applications in image pattern discovery and multiple-feature clustering • Can be applied to other general clustering problems