大数据分析笔记3-聚类

[SSK015]

https://ssk015.github.io/dashuju3/

Click here Slide(pdf)

What is Clustering?
That is to say, given a set of points, We can define a concept of distance between these points. Then, we group the points into some number of clusters, which is known as “簇” in Chinese.

distance : mainly Euclidean or Jaccard

Why is Clustering hard?
Too many dimensions: isolated points

Two main methods:
Hierarchical: bottom up and top down
Assign: assign points to a Existing cluster

Hierarchical(between clusters):
note:

represent a cluster.
determine the nearness of clusters.
when to stop merging clusters.
Euclidean case:

centroid(average)
(1) distance of centroids
(2) shortest distance between two clusters

UnEuclidean case:

Approach 1

choose Clustroid(a exisiting point)

maxium/average/square of dis

various distance and cohesion measures
Approach 2

the collection of points.
define inter-cluster distance.
min of two or avg of all pairs.
Approach 3

the collection of points.
define a notion of cohesion, merge similiar unions.
diamter, avg dis, density
3.
design 1: convex clusters
design 2: concentric clusters.

K-Means(Assignment)

definition: a method, not a algorithm.

method: Before convergence(all the points don’t move), assign points and update centroids

select k: try different k, get the value when avg dis to centroid stop changing dramatically.