Entry 9: Train Model
At the end of the Entry 8 noteboook I had a standardized dataset. Now it’s time to try training a model.
The notebook where I did my code for this entry can be found on my github page in the Entry 9 notebook
The Problem
Train a model using the standardized dataset where the collinear features have been removed.
The Options
Scikit-Learn is the most common package for building models in Python. Implementing by hand is also an option. Andrew Ng teaches manual implementation of linear and logistic regression in his Machine Learning course on Coursera. However, using a package will allow me to iterate through multiple types of models faster than implementing and optimizing each one of them by hand. As an added benefit, by using an open source package there are a lot of people contributing to it, so the code and mathematical optimizations are probably better than anything I could come up with without several more years of experience under my belt.
The Proposed Solution
As discussed in Entry 2, Scikit-Learn lists 17 different kinds of algorithms. While working through the process on this first model, I’m going to stick to one of the most basic models - linear regression.
The actual implementation of the model training went smoothly.
- The data was all ready after Entry 8, so no further pre-processing was required.
- Separating the target from the attributes was easily accomplished with pandas.
- Training the model was as easy as specifying the model type and calling the
fit()method.
The Fail
There were several pitfalls in this exercise.
Missing Values
While there were no missing values in this particular, hand-curated dataset, missing values are extremely common in the wild. I’ll need to address this for any kind of standardized process or automated pipeline.
Validating the model
This dataset is really small (only 11 observations), so splitting it into train and test sets would leave very little data to work with or to validate on. I tried looking at the coefficients and score, but these values are currently meaningless to me.
I decided to skip this step for now. A full set of model diagnostics will be developed in the next series of entries based on week 6 of Andrew Ng’s Machine Learning course, the books in my collection, and supplementary online resources.
Retain Scaling
I had planned on making a prediction in this entry too. Unfortunately, I forgot about retaining scaling parameters during my pre-processing stage. For this particular dataset, I could get around the categorical encoding by copying the line of interest from the already encoded version of the data (I just made two copies of Mars’ row then changed the surface pressure to the high and low values that humans can withstand based on survivable conditions (nitrogen narcosis in divers being the high end and conditions experienced by climbers on Mt Everest being the low end), but the categorical encoding will need to be retained for any other dataset I’d want to evaluate (I wouldn’t be copying existing data - it should be data that the model has never seen before).
Process Order
Separating the target variable after all the pre-processing leaves me with a scaled value. The scaled value has no meaning and would need to be un-scaled to get the actual prediction. In reviewing the purpose of scaling (to allow the model to quickly converge on a better solution), there really isn’t a reason to scale the target value. As such, it should be separated from the features before pre-processing.
Up Next
Fixing Entry 9 failures and making a prediction)