LFD Book Forum

LFD Book Forum (http://book.caltech.edu/bookforum/index.php)
-   Homework 7 (http://book.caltech.edu/bookforum/forumdisplay.php?f=136)
-   -   REQUEST: Q7 How-To (http://book.caltech.edu/bookforum/showthread.php?t=4654)

galo 01-27-2016 11:50 PM

Can anybody share the rundown to get the answer? I can't grasp a way to do it. Plus it gets frustrating because I feel like it's more of an algebra problem rather than a Machine Learning one.

tddevlin 08-14-2016 03:56 PM

Re: REQUEST: Q7 How-To
**Spoiler Alert: This post contains the full solution**

First let's make sure we have the right picture.


So P_1 and P_2 are sitting on the x-axis, while P_3 is somewhere to the right of the y-axis at height 1. For this dataset, leave-one-out validation entails fitting our model to two of the points, then testing the fit on the third. Let's start with the constant model, h_0(x) = b. When we fit this model on two data points, b will simply be the average of the y-coordinates of the two points.
  • Leaving P_1 out, we find b = 1/2. The error is e_1 = (h_0(x_1) - y_1)^2 = (b - 0)^2 = (1/2)^2.
  • Leaving P_2 out, we also find b = 1/2. Again, e_2 = (h_0(x_2) - y_2)^2 = (b - 0)^2 = (1/2)^2.
  • Finally, leaving P_3 out, b = 0. The error is e_3 = (h_0(x_3) - y_3)^2 = (0 - 1)^2 = 1.

The overall cross-validation error is the average of the three individual errors, E_{\text{val}}(h_0) = \frac{e_1 + e_2 + e_3}{3} = 1/2, as you can verify. Looking ahead, we would like to find the value of \rho that makes E_{\text{val}}(h_1) = 1/2.

Let's turn to the linear model, h_1(x) = ax + b. The easy case is when P_3 is left out. The resulting fitted line is simply y = 0 and the error is e_3 = 1.

Things get more complicated when P_2 is left out. We need to find the equation of the line through P_1 and P_3. Using slope-intercept form and rearranging, you can check that the fitted line has slope equal to its intercept, a = b = \frac{1}{1 + \rho}. The error on P_2 is e_2 = (h_1(x_2) - y_2)^2 = (ax_2 + b - 0)^2 = \left( \frac{2}{1+\rho} \right)^2.

A similar derivation yields e_3 =  \left( \frac{-2}{\rho - 1} \right)^2.

Putting it all together gives us E_{\text{val}}(h_1) = \frac{1}{3} \left[ 1 + \left( \frac{2}{1+\rho} \right)^2+ \left( \frac{-2}{\rho - 1} \right)^2 \right]. If we set this equal to 1/2 (the error from the constant model), we have a quadratic equation in one unknown, which we can solve using the quadratic formula (alternatively, dumping the whole equation into WolframAlpha gives you the roots directly).

Hope that helped!

All times are GMT -7. The time now is 02:18 PM.

Powered by vBulletin® Version 3.8.3
Copyright ©2000 - 2019, Jelsoft Enterprises Ltd.
The contents of this forum are to be used ONLY by readers of the Learning From Data book by Yaser S. Abu-Mostafa, Malik Magdon-Ismail, and Hsuan-Tien Lin, and participants in the Learning From Data MOOC by Yaser S. Abu-Mostafa. No part of these contents is to be communicated or made accessible to ANY other person or entity.