A constrained growing grid neural clustering model

摘要

A novel extension of the growing grid (GG) algorithm is proposed in this paper. The learning behavior of the traditional GG model is affected by two factors i.e. similarity between output units and their associated input vectors, and lateral connections between output units. Based on the assumption that the more active unit should move more towards its associated input vector, the constrained GG emphasizes the effect of the lateral connections between output units in a grid, and neutralizes the effect on the distance between the input vector and neighbors of the best matching unit (BMU). A constrained learning rule is tested by fifteen data sets, i.e. the square, animal, iris, ionosphere, sentiment polarity data sets and the fundamental clustering problem suite (FCPS), which includes ten data sets. Based on five evaluation measures, i.e. average quantization error, error entropy, BMU activation rate, average map scope and topographic error, the performance of GG is improved if the constrained learning rule is used. In addition, we use the t-test to test whether or not the proposed models outperform the traditional GG significantly. Except in some cases, the experiments conclude that the proposed approach can significantly improve the traditional GG model based on five evaluation criteria for fifteen data sets.