Stanford ML - Lecture 9 - Clustering

1. Unsupervised learning introduction

• supervised learning - training set: 
• unpervised learning - training set: 

2. K-means algorithm

• randomly select K cluster centroids
• repeat
  • for every example, select the closest centroid to it
  • update the new centroid

3. Optimization objective

4. Random initialization

• randomly initialize K-means algorithm many times, e.g. 100 times, compare the optimization objective function, pick the lowest cost

5. Choosing the number of clusters

• elbow method
  • select the break point

有监督学习最常见的是regression & classification

无监督学习包括density estimation & clustering

• PCA和很多deep learning算法都属于无监督学习

回归问题中有可能因为学习率设置过大产生随着迭代次数增加，cost function反倒增大的情况。但聚类是不会产生这样的问题的，因为每一次聚类都保证了使J下降，且无学习率做参数。

Reference: http://blog.youkuaiyun.com/abcjennifer/article/details/7914952