最近要在spark上做一個聚類的項目，數據量和類的個數都比較大。KMeans效果尚可，但是有點慢，因而重新看了下常用的算灋。最終選用mini-batch kmeans，使用類似kmeans++的方法來初始化類中心。

kMeans   attention: init centers（randomize vs kMeans++）

mini-batch kMeans   loops: random select samples；find closest for all samples；update centers for each sample

mean shift   init: get centers by bandwidth   loops: find neighbors of centers；update centers；de-duplicate

DBSCAN   init: get densest core samples   loops: get more core samples nearby old samples

Ward hierarchical   init: each sample as center   loops: merge to minimize RMSE within clusters

Spectral clustering   steps: similarity matrix S；S=UV；kmeans of U

AP cluster   init: get S；Rik=0，Aik=0   loops: Rik = Sik - Max_k'！=k（Aik' + Sik'）；Aik = min（0，Rkk + Sum_i'！=i，k Max（0，Ri'k））；Akk = Sum_i'！=k Max（0，Ri’k）   end: for any i，Max_k Rik + Aik as it’s exemplar

Topic Model（LDA）

其實scikit-learn實現了很多算灋，也有現成的數据集可以做做實驗。例如：http://scikit-learn.org/stable/modules/clustering.html 上有一些效果圖，和算灋擴展性的說明。

                       A comparison of the clustering algorithms in scikit-learn

發佈了19篇原創文章·獲贊16·訪問量23萬+