Rethinking of K-means

FCM方向做多了，回头顺便想了想K-means的缺陷:

1. With respect to the object function, standard K-means clustering algorithm shows underlying sentisitity to outlier and noisy datasets.
2. Face with the size-imbalance & dimensional-imbalance problems.
3. Donnot meet the real-time requirements for large scale datasets clustering.
4. To deal with high-dimenstional datasets, additional low-rank processings or sparse weighting operators for feature selection are required.
5. Hard Logic rooting in K-means is not quite suitable for partitioning datums of overlapped distributions or lying in manifold space, resulting in the degradation of model generalization.
6. Etc.