Doing It Wrong

HAL8999 7/100

I was sick yesterday but did spend some time looking over some “cheat sheets” that people had put together for various machine learning topics. Some were good, some were just stupid (I’m looking at you Machine Learning in Emoji). Also went through a very simply classifier based on the iris data set.

Model Selection	Microsoft Azure
Neural networks	Neural Network Graphs
Python 4 Big Data	Python 4 Data Science
Scikit-Learn

HAL8999 6/100

• Watched the “Learn how to Learn” Google talk on youtube
• Updated the jupyter notebooks for handson-ml from github and read through the Ch2 notebook to address the CategoricalEncoder issue from yesterday
• Looked at a basic transformer

Part of building a data pipeline is likely to include the creation of custom transformer classes to perform operations specific to the project or data source. For example, one of the products I work on stores xml data in a database with the newlines encoded as ‘\n’. When the data is pulled from the database those ‘\n’ sequences are converted to newline characters before the xml is passed to the parser. It’s a very simple operation but without it the data would fail xml validation.

The scikit-learn package provides a structure for building transformers for a data pipeline that is based on duck typing i.e. “looks like a duck, walks like a duck, etc” rather then through object inheritance. Essentially, if your class has fit(X) and transform(X) methods, it counts as a transformer.

HAL8999 – 5/100

Today while going back through the Hands On Machine Learning book Ch2 I learned that the CategoricalEncoder referenced in the section on handling categorical attributes still isn’t in scikit-learn. I checked the reqirements.txt which shows scikit-learn=0.19.1. Checking my virtualenv, I should be good.

Turns out that the CategoricalEncoder isn’t going to be in scikit-learn until 0.20 so to get it you have to grab 0.20 from Github rather than just use pip.

Fucking hell… 

So, if you’re going to write a book, it’s probably a good idea to use the stable branch of your libraries rather than the bleeding edge dev branch.

It will be a good exercise to convert the book’s example code to work with the standard OneHotEncoder but I’ve always been a fan of “just works” as a design principle.

HAL8999 – [3,4]/100

• Chapter 2 of Hands on ML continues
• Creation of test sets
• Stratified sampling
• sklearn’s StratifiedShuffleSplit
• Visualizing data with matplotlib
• Coorelation coefficients

Getting a good train-test split

Since you can’t train a model and just expect it to work well right out of the box it’s standard practice to split off about 20% of the data set to test the model against. The naieve way to do this is to just grab 20% of the data at random but that runs into a number of issues:

• depending on how you do it, you may grab different train/test sets every time the model runs
• grabbing data points at random can result in sampling bias to creep in if you happen to get an unrepresentative sample

Solution?

Stratified sampling

Rather than just grabbing data points at random we can ensure that we can get a more representative distribution of sampled data points for some attributes (sex, income, ethnic background, etc) to ensure that random selection hasn’t introduced bias into the training and test sets.

In this example we can be pretty certain that median income correlates strongly with median housing price and we want to be certain we get a representative distribution of districts with respect to median income. The way to do this is to add a column to the data set that groups median income into categories and we can then sample based on the category. This will improve our chance of getting a more representative sampling of the underlying median income attribute.

Once the test set has been selected out we work entirely with the training set so as to not introduce bias based on knowledge of the test data.

HAL8999 – 3/100

• Chapter 2 of Hands on ML
  • Cost functions
  • virtualenv setup
  • code to get the dataset

The chapter follows a rudimentary machine learning project from business case to final product. California census data is analyzed to build a model which will predict media housing price in a district based on other factors using a linear regression model with a Root Mean Square Error (RMSE) function to measure performance i.e. as a cost function.

The function h is the “hypothesis” function which operates on the feature vector \(x^{(i)}\). RMSE isn’t the only cost function by any stretch of the imagination but it seems to get a lot of use.

From this point the author goes through the dev environment setup process I went through a few days ago and it’s pretty clear from the instructions that the work is being done on a Mac. 

The code to download the housing tarball is a little sloppy and would have downloaded it every time I ran the cell so I added a simple test to only do the download if the file didn’t already exist.

Achievement: HAL 8999 – 2/100

• Completed chapter 1 of Hands On ML and worked through the exercises
• Modified yesterdays example to also do both k-nearest-neighbors with both three and four neighbors. Four neighbors was further from the linear regression than three demonstrating that more is not always better.

Short list but Ch1 is something of an overview so a lot of concepts get thrown in with not a lot of context or depth of discussion so I found I got to the end and had a hard time connecting what I’d read with the specific questions asked at the end of the chapter. I ended up paging back through the chapter to locate the answers to questions which were oddly specific as opposed to focusing on the broad underlying concepts.

The fact that I did a chunk of the reading while grilling tri-tip and elote and then later when in the post steak and mexican corn food coma might also be part of why and ended up paging back through the chapter so much.