Saturday, 28 February 2015

Review 3.2: Find Independant Features - Use of NMF(Non-Negative Matrix Factorization)

There are two examples of using NMF. Because NMF is a unsupervised learning technology. So it acts a little bit like clustering. 

The first example is assigning themes to the acticles from different news websites. We can download those article first, and then construct a matrix in the following format:
|5,0,4|
|3,4,5|
|1,2,9|
The rows are article titles, and the columns are different words. So the matrix shows the word count of different word in each article. The we can use NMF to factorize the matrix to a weight matrix and a feature matrix. Weight matrix is composed with article titles and features , besides, feature matrix is compose with features and words. So according the these two matrix , we can find out the relationship between words count and articles, what we always named as themes.

The second example uses NMF to mine information in the stock market. The stock market data can be download from any financial site. Then rows can be represented as dates, columns represented as different stock. The data can be close volumns. Once, the matrix is factorized, just like the first example, we can find out the relationship between stock price and date. For instance, google's trading volumns is increased a lot on 20 Jan,06 because google announced it would not given information about its search engine usage to the government.

No comments:

Post a Comment