DAAL库K-means算法空簇处理机制咨询——对比MATLAB三种方式

Having worked extensively with Intel DAAL's machine learning libraries, I can clarify how its K-means implementation deals with empty clusters during optimization—since this detail is easy to miss in the official docs:

• Default core behavior: Unlike MATLAB’s configurable flags, DAAL’s K-means doesn’t let you pick between erroring out, removing clusters, or reinitializing. Instead, it automatically maintains the specified number of clusters (k) by reinitializing any empty cluster’s centroid. The strategy here mirrors MATLAB’s third option: it selects the single observation that’s farthest from its currently assigned centroid (or more precisely, the farthest from the closest non-empty cluster’s centroid) and sets that point as the new centroid for the empty cluster.

• No explicit configuration switch: DAAL doesn’t expose a direct parameter to change this behavior. If you need alternative handling—like treating empty clusters as an error or removing them entirely—you’ll have to implement post-processing logic: run the K-means algorithm, then check the cluster assignment results to filter out clusters with zero assigned samples, or add validation steps before/after iterations to flag empty clusters as errors.

内容的提问来源于stack exchange，提问作者BayesianMonk