使用 Mahout 进行 K 均值聚类
我正在使用此处给出的聚类技术 用于对大型数据集进行聚类，这在 Mahout 示例中给出。然而，当我可视化特定的聚类时，我得到下图。
我真的很难理解这实际上意味着什么，并且有几个问题。
- 所有彩色线表示什么？
- 这么多簇是什么意思？
- 为什么只有少数区域拥挤，而其他区域却不拥挤？
- 为什么很少有彩色线相互重叠？
I'm using the clustering technique given here for clustering a large dataset, which is given in Mahout examples. However, when I visualize the particular clustering I get the following figure.
I'm really struggling to understand what this actually means and have several questions.
- What does all the coloured lines indicate?
- What does so many clusters mean?
- Why are few areas crowded, and why aren't the other areas crowded?
- Why are few colored lines overlapping each other?
如果你对这篇内容有疑问，欢迎到本站社区发帖提问 参与讨论，获取更多帮助，或者扫码二维码加入 Web 技术交流群。
绑定邮箱获取回复消息
由于您还没有绑定你的真实邮箱，如果其他用户或者作者回复了您的评论，将不能在第一时间通知您！
发布评论
评论（1）
k-means 并不是最先进的聚类技术。圆圈作为一种可视化技术具有误导性，它实际上将数据空间划分为 Voronoi 单元（在维基百科上查找）。它还更喜欢大小相似的集群。
我假设不同的颜色表示 k 均值的不同迭代。它需要多次运行来优化其结果（通常只能达到局部最小值，并且不同的运行将导致不同的结果）。所以我猜结果还不是很稳定。它们移动缓慢，这就是它们重叠不多的原因。
簇的数量是 k-means 的一个参数。它通常表示为
k。 k-means 无法确定簇的数量，但如果使用多个 k 值运行它，您可以测试哪个结果最适合数据集。k-means 不考虑密度。为此，您需要一种基于密度的聚类算法。 k-means 更喜欢大小相似的簇。你的“k”可能太高了。
由于它们是迭代更新的，因此不同的迭代不应重叠太多。
k-means is not the most advanced clustering technique. Circles as a visualization technique are misleading, it's actually partitioning the data space into Voronoi cells (look it up on Wikipedia). It also prefers similar-sized clusters.
I assume that the different colors indicate the different iterations of k-means. It requires several runs to optimize its result (which usually only reaches a local minimum, and different runs will result in different results). So the results aren't very stable yet, I guess. They shift only slowly, which is why they don't overlap much.
The number of clusters is a parameter for k-means. It's commonly denoted as
k. k-means cannot determine the number of clusters, but you can test which result fits the data set best, if you run it with multiple values of k.k-means doesn't look at density. You need a density-based clustering algorithm for that. k-means prefers similar-sized clusters. Your "k" is probably too high.
Since they are iteratively updated, the different iterations shouldn't overlap much.