Self-organizing map with distance measure defined by data distribution

In this paper, a new distance measure for learning improvement of self-organizing map is described. The new distance measure is defined based on data distribution, thus it can be efficiently used for real application, in which data is often distributed in nonlinear manifold. The authors have reported graph-based distance, but it requires high computation performance. To reduce computational cost, we define energy function in data space, and distance is calculated using energy function. Experimental results using simple data show effectiveness of the proposed method.