For a slightly more complex example, let's look at another common machine learning application, such as predicting if a patient is at risk for a certain disease. In this case, we want a simple yes or no of is a patient high risk for heart disease?

Notice how we set a simple yes or no. We could also say, what is their risk score? But in this case, we're asking for a categorical high risk or not high risk, rather than a score of 0 to 100% likelihood of developing heart disease.

So this shows you how the same type of problem, depending on the business requirement of yes, no, or a score can change the underlying engineering and data science approach.

In more complex issues, you'll often have a blend of lots of different types of features in order to predict your outcome, aka your label. And here, we have a good mix of categorical features such as gender, are they a smoker, and numeric features such as weight, cholesterol, and age.

This blend can make it a little overwhelming to start on machine learning projects so feel free to disregard some of them, but know that your model accuracy will suffer. Once again, don't be afraid to iterate in order to start building your understanding of the problem.

To put these features in a more structured format, you could see here we have the features on the left, gender, age, weight, cholesterol, and blood pressure, and the label on the right. Notice how we're saying, does the patient have heart disease? It is a categorical label. We're not measuring heart health or other factors. It's simply, do they have a heart disease diagnosis? And it is a binary decision.

To start again with our flow chart, we have a good business understanding. We want to know yes or no, is this patient high risk for heart disease? We have a labeled training set, which means we can go with supervised machine learning. And due to the business requirements, we have a binary classification. Do they or don't they have a high risk for heart disease?

Now remember, we might have to go with multi-classification or regression for different types of requirements from the business, but in this example, it's yes or no, so it's a binary classification problem. And finally, in terms of algorithm, we don't have a clear linear relationship to associate this with.

So in this case, we need to start thinking about different types of algorithms. K-nearest neighbor or KNN is an index-based algorithm that uses non-parametric methods for classification. Simply put, this means that the data is plotted and the distance between different data points on multiple axes is used to determine the quote K points that are closest to the sample point.

Basically, we're trying to predict how far away any test or production data is from the training data, and then simply associating it with that data point. K-nearest neighbor algorithms can get pretty complex quite quickly. So we're gonna avoid going into too much depth here, but it's very important to see K. This is what's called a hyperparameter, and it is a setting that fundamentally controls how the algorithm works. 

Basically, by setting K equal to one or more than one, we control how much it values its nearest neighbors and how it handles clustering. This is often set by the data scientists themselves or with some of the more modern programs you have what's called hyperparameter tuning and the system will automatically try multiple values of K and report back to you which model has the best fit.

Just know that some machine learning models, in addition to training, need to have specific types of tuning through what's called hyperparameters. That's a little beyond the scope of this class. It's really higher level, level three and level four stuff to know how those hyperparameters play exact roles in machine learning.

Lectures

Course Introduction - Explaining Concepts - Models - Understanding Training Data Sets - How to Choose? - Case Study: Home Prices - Case Study: Animal Classification - Case Study: Targeted Marketing