homework_03

Due Mar 4th, 6pm

Code To receive full credit all unit tests must pass and one copy of the exercises must be completed.

Exercises: 6.4.2, 6.4.3, 6.6.1, 6.7, 6.9.1, 6.12.1, 6.12.2, 6.14.1, 6.19.1


To start

Clone the repo into a local directory named homework_03. Do not use the original repo name. Replace X below with your team name.

git clone https://github.com/columbia-applied-data-science/homework_03_team_X.git \
  homework_03

See demo.py and the tests to get an idea of how things work.


Numerical techniques

See this section in the lecture notes information about pseudo inverses.


5-fold cross validation

You will make a 5-fold cross validation module. This is used as a way to pick out your regularization parameter delta. Our 5-fold cross validation is:

For every delta:

  1. Divide the data up into 5 equal chunks
  2. Pick out the first chunk as a cross-validation set, and group the other 4 together as training data.
  3. Fit the model using the training data and use the cross validation set to measure both the training and cross-validation squared error |Xw - Y|^2
  4. Repeat 5 times, each time using a different chunk as the cross validation set.
  5. Average the training and cross-validation errors across the 5 folds.

Compare the average cross-validation errors and use this to choose delta. Note that the training error should not be used to choose delta. It is there to serve as a reality check and to diagnose the degree of over/under fitting.


Caution!

These routines are very picky about array shape. Some functions, e.g. np.dot, return arrays who have shape = (N,) (a tuple with only one element). In that case, you will often have to reshape this into a proper two dimensional array. The docstring for linear_reg.fit() tells you when to do this.

Two functions, linearreg.fit() and crossvalidator.cross_val() can handle pandas objects as their input. The others may or may not. However, these are the only public methods in their modules, so this is ok.



blog comments powered by Disqus

Published

19 February 2013

Tags

Subscribe to RSS Feed