PCA

PCA

Principal Component Analysis is a method that analysis data, it is used for dimensional reduction, data visualising and more. The sole purpose of PCA is to find the principal components that describes the most variation in the data another way of saying this is PCA represent the data in new axes in a coordinate system.

If we have a data-set of cells, we don't know what type of cells they are, but we know how many types of cells there is. In this example we give them some sugar. Each column describes how much a cell grows. If we only have asked two cells it is easy to visualize in a graph. If we have three cells, we could either visualize it with a 3d-graph or draw three 2d-graph to see if there is any correlation. If there is more than 3 cells it becomes more difficult to do by plotting the data into one graph, and the amount of 2d-graph that shows all the data would be too much to look at. This is where PCA becomes handy. To do PCA there are some steps to follow.

Step 1 Normalising

---

Here we find the average of the column and subtract it from the original column.

$\bar{x} = \frac{1}{N}\Sigma_{n=1}^N x_n$

N is the number of columns.

Step 2 Covariance matrix

---

Here we calculate the Covariance matrix with the formula

$S=1/N(X-\bar{X})(X-\bar{X})^T$

This finds the ratio between the two sets.

Step 3 Eigenvectors and principal components

---

We calculate the eigenvectors and eigenvalues for the Covariance matrix. Where the eigenvectors is $u_1, ... u_N$ and the eigenvalues is $\lambda_1, ..., \lambda_N$. This is found by the formula

$Su_i=\lambda_i u_i i\in{1,2,...,N}$

This step is calculated multiple times, to find the best suitable eigenvectors.

Step 4 Chose M components

---

Sort the eigenvectors in decreasing order, and chose your number M, put the chosen data into a matrix as $M < D$.

Note if we don't want to reduce the data this step is unnecessary

Step 5 Transform data

---

We can now transform the data into the linear combinations we made by $P=(X-\bar{X})^TU_{1:M}$

If the data should not be reduced we use N instead of M.

Further understanding

---

If the description this video goes over PCA but uses a more grafic approach on PCA

https://www.youtube.com/watch?v=FgakZw6K1QQ

Sources

---

Christopher M. Bishop. Pattern Recognition and Machine Learning. Information Science and Statistics. New York: Springer, 2006. 738 pp. isbn: 978-0-387-31073-2.