aiacademy: 機器學習 Unsupervised Learning

Unsupervised Learning

• Outline

  • Unsupervised learning

  • Dimension reduction

    • Principal component analysis (PCA)
    • T-Distributed Stochastic Neighbor Embedding (t-SNE)

  • Clustering

    • K-means
    • Hierarchical clustering

Dimension Reduction

跟 Coursera 的 Data Compression 一樣

• Why dimension reduction ?

  • compress data preserve useful information
  • Data visulization

Principlal Component analysis (PCA)

先作筆記，等看到　coursera 的時候再來複習！！！

• PCA
  • 把高維度的點，投影到低維度上面，且希望在低維度空間中保有在高維度中的性質！

PCA in general

• Perform PCA by computing the eigenvectors of the k largest eigenvalues of the covariance matrix

PCA on MNIST

T-SNE (T-distributed Stochastic Neighbor Embedding)

• Goal: find locations in low dimensions such that the distance between points are preserved

• T-SNE allows non-linear transforms from the original data point to the new data point

T-SNE

Why different silmilarity measures?

• Crowding problem: distance between distant data points in the low dimension will not be large enough by SNE, which uses the same similarity measure for both low and high dimensional space

T-SNE on MNIST

Summary

• Both PCA and t-SNE project data points into low dimension

  • PCA allows only linear projection
  • T-SNE allows non-linear projection

• Adv of PCA

  • Interpretability
  • Can project new data points

• Adv of t-SNE

  • Visualization

Clustering

Hierarchical clustering

• Types of hierarchical clustering

  • Agglomerative (bottom-up)
    • Start with each data point as a cluster
    • Merge two closest clusters until only one cluster left
  • Divisive (top-down)
    • Start with one cluster
    • Each step split a cluster until each cluster contains one data point

Agglomerateive example

Different cluster result

Summary

• Clustering
  • k-means and hierachical clustering