Abstract:

Based on the quotient space granularity theory，a multi-granularity selflearning clustering algorithm (MSCA) is presented for problems with non-uniform cluster density. By constructing a feature clustering tree and defining a granularity function，MSCA solves problems layer by layer and learns clustering granularity dynamically by itself in each step. Traditional clustering algorithms with global parameters cannot discover data features in various layers，and their parameters depend on professional experience seriously，while MSCA can overcome  these defects. Both theory analysis and experimental results show that MSCA can discover key clustering layers and clusters with arbitrary shape. Furthermore，it is insensitive to noise and has a satisfactory computing complexity.