Principal Component Analysis

Video material (part 1) Download video: Link (plays in Google Chrome) [290.1 Mb]
Video timing
00:00 to 21:37 Recap and overview of this class 21:38 to 42:01 Preprocessing: centering and scaling 42:02 to 57:07 Geometric view of PCA
Video material(part 2) Download video: Link (plays in Google Chrome) [305.5 Mb]
Video timing
00:00 to 58:07 Details coming soon
Video material (part 3) Download video: Link (plays in Google Chrome) [293.7 Mb]
Video timing
00:00 to 1:03:39 Details coming soon

Class notes

<pdfreflow> class_date = 16 September 2011 [1.65 Mb] button_label = Create my projector slides! show_page_layout = 1 show_frame_option = 1 pdf_file = lvm-class-2.pdf </pdfreflow>

• Also download these 3 CSV files and bring them on your computer:
  • Peas dataset: http://datasets.connectmv.com/info/peas
  • Food texture dataset: http://datasets.connectmv.com/info/food-texture
  • Food consumption dataset: http://datasets.connectmv.com/info/food-consumption

Class preparation

Class 2 (16 September)

• Reading for class 2
• Linear algebra topics you should be familiar with before class 2:
  • matrix multiplication
  • that matrix multiplication of a vector by a matrix is a transformation from one coordinate system to another (we will review this in class)
  • linear combinations (read the first section of that website: we will review this in class)
  • the dot product of 2 vectors, and that they are related by the cosine of the angle between them (see the geometric interpretation section)

Class 3 (23 September)

• Least squares:
  • what is the objective function of least squares
  • how to calculate the two regression coefficients \(b_0\) and \(b_1\) for \(y = b_0 + b_1x + e\)
  • understand that the residuals in least squares are orthogonal to \(x\)

• Some optimization theory:
  • how an optimization problem is written with equality constraints
  • the Lagrange multiplier principle for solving simple, equality constrained optimization problems

• Reading on cross validation

Update

This illustration should help better explain what I trying to get across in class 2B

\(p_1\) and \(p_2\) are the unit vectors for components 1 and 2. \( \mathbf{x}_i \) is a row of data from matrix \( \mathbf{X}\). \(\hat{\mathbf{x}}_{i,1} = t_{i,1}p_1\) = the best prediction of \( \mathbf{x}_i \) using only the first component. \(\hat{\mathbf{x}}_{i,2} = t_{i,2}p_2\) = the improvement we add after the first component to better predict \( \mathbf{x}_i \). \(\hat{\mathbf{x}}_{i} = \hat{\mathbf{x}}_{i,1} + \hat{\mathbf{x}}_{i,2} \) = is the total prediction of \( \mathbf{x}_i \) using 2 components and is the open blue point lying on the plane defined by \(p_1\) and \(p_2\). Notice that this is just the vector summation of \( \hat{\mathbf{x}}_{i,1}\) and \( \hat{\mathbf{x}}_{i,2}\). \(\mathbf{e}_{i,2} \) = is the prediction error vector because the prediction \(\hat{\mathbf{x}}_{i} \) is not exact: the data point \( \mathbf{x}_i \) lies above the plane defined by \(p_1\) and \(p_2\). This \(e_{i,2} \) is the residual distance after using 2 components. \( \mathbf{x}_i = \hat{\mathbf{x}}_{i} + \mathbf{e}_{i,2} \) is also a vector summation and shows how \( \mathbf{x}_i \) is broken down into two parts: \(\hat{\mathbf{x}}_{i} \) is a vector on the plane, while \( \mathbf{e}_{i,2} \) is the vector perpendicular to the plane.