加权闵可夫斯基K-Means的指数选取策略

摘要

与传统 K-Means 相比，加权闵可夫斯基 K-Means(MWK-Means)需要自适应获取特征权重并选择合适的闵可夫斯基指数。无监督选取指数策略是计算每个指数的三种尺度值，根据三种尺度的选取标准得到各自最好的指数，然后选取较接近的两个指数求均值。在这种策略的启发下，提出了基于排名的闵可夫斯基指数选取策略，将三种尺度的值分别进行排名，每个指数通过选取两个较接近的排名相加得到综合排名来确定指数。用这两种指数选取策略分别对UCI数据集进行实验，结果表明，基于排名的选取策略较优。%Compared to the traditional K-Means, the MWK-Means needs to obtain feature weights adaptively and select suitable exponent. Choosing the Minkowski exponent in an unsupervised setting is a way to calculate three-scale values of each exponent. It gets the best of each exponent based on the selection criteria of three scales, and then gets the mean of two closer exponents. According to this strategy, we put forward a new strategy of selecting the Minkowski exponent based on ranking, ranked the values of three scales each. Then, we added the two closer rankings of each exponent as comprehensive rank and used them to determine the final exponent. This paper used the above two strategies of selecting Minkowski exponent to test UCI dataset. The result shows that the new strategy is better.