Data Clustering: Principal Components, Hopfield and Self-Aggregation Networks

We present a coherent framework for data clustering. Starting with a Hopfield network, we show the solutions for several well-motivated clustering objective functions are principal components. For MinMaxCut objectives motivated for ensuring cluster balance, the solutions are the nonlinearly scaled principal components. Using scaled PC A, we generalize to multi-way clustering, constructing a self-aggregation network, where connection weights between different clusters are automatically suppressed while connection weights within same clusters are automatically enhanced.

• Type: Conference
• Year: 2003
• Where: IJCAI
• Authors: Chris H. Q. Ding
• Keywords: Clustering Objective Functions, Connection Weights, IJCAI 2003, IJCAI 2007, Principal Components