Model training and evaluation

As mentioned previously, we'll be predicting customer satisfaction. The data is based on a former online competition. I've taken the training portion of the data and cleaned it up for our use. 

This is an excellent dataset for a classification problem for many reasons. Like so much customer data, it's very messy— especially before I removed a bunch of useless features (there was something like four dozen zero variance features). As discussed in the prior two chapters, I addressed missing values, linear dependencies, and highly correlated pairs. I also found the feature names lengthy and useless, so I coded them V1 through V142. The resulting data deals with what's usually a difficult thing to measure: satisfaction. Because of proprietary methods, no description or definition of satisfaction is given.

Having worked previously in the world of banking, I can assure you that it's a somewhat challenging proposition and fraught with measurement error. As such, there's quite a bit of noise relative to the signal and you can expect model performance to be rather poor. Also, the outcome of interest, customer dissatisfaction, is relatively rare when compared to customers not dissatisfied. The classic problem is that you end up with quite a few false positives when trying to classify the minority labels.

As always, you can find the data on GitHub: https://github.com/PacktPublishing/Advanced-Machine-Learning-with-R/blob/master/Data/santander_prepd.RData.

So, let's start by first loading the data and training a logistic regression algorithm.