-->

Scalable and Distributed Clustering via Lightweight Coresets

Posted on By Marquis

来自:苏黎世联邦理工学院,名校啊

终于看到和DNN无关的clustering文章了,没有那么多detail的烦恼,好激动。
最近几天一直陷到DNN的文献里边了,心情不爽啊,赶紧搞个纯聚类的压压惊。

啊啊,看不懂啊
我看到了各种bound……
看到第三页就没法看了……

搜了一下,发现作者主页,啊就是NIPS16 拿到oral的那个!搞出来一个 Fast and Provably Good Seedings for k-Means

算了先看实验吧
看到人家用了四个数据集,exicited!

  1. KDD — 145’751 samples with 74 features measuring the match between a protein and a native sequence
    (KDD Cup, 2004).
  2. CSN — 7GB of cellphone accelerometer data processed into 80’000 observations and 17 features
    (Faulkner et al., 2011).
  3. MSYP — 90 features from 515’345 songs of the Million Song datasets used for predicting the year of songs
    (Bertin-Mahieux et al., 2011).
  4. CODRNA — 8 features from 488’565 RNA input sequence pairs (Uzilov et al., 2006).

人家用了1.5TB memory,上G的数据集才明显看出来效果(里头用了抽样),看样子一个好的结果的确得有各种付出和机遇。

文章的主体其实在第三章就有了,就是 Definition 1 (Lightweight coreset for k-Means) ,形式很简单,后边第四、五章貌似就是为这一章服务的。