Building algorithms that can find or “learn” relationships in data is the broad definition of machine learning; supervised learning is the branch of machine learning that focuses on identifying relationships between previously measured data properties.
Finding connections between previously measured and unmeasured variables is the idea behind unsupervised learning, the other type of machine learning.
A common supervised learning problem that you may face in the real world: determining the correlation between a home’s attributes and its worth. It is obvious that there is some correlation between a home’s desirability as a place to live or own and features like the number of rooms, square footage, or closeness to schools. Since we have already measured these attributes, the overall goal of supervised learning is to find these associations.
By “measure,” we mean that every attribute has been properly specified and quantified. This part of the problem would be much more difficult if we had other different types of information, like natural language descriptions of the house’s neighborhood, as many aspects of a house, like the number of bedrooms, square footage, and so on, naturally lend themselves to being represented as numbers.
After converting our “characteristics” into numerical values, we must choose the appropriate representational framework. Representing each set of numbers for a single observation, like a single house, as a row of data and then stacking these rows on top of one another to create “batches” of data that will be fed into our models as two-dimensional ndarrays is one that is almost universal across machine learning and turns out to make computations simple. The predictions will then be returned by our models as output ndarrays, one for each observation in the batch, each forecast in a row, similarly piled on top of the others.
Here are some definitions: we state that the number of features in our data is equal to the length of each row in our ndarray. A characteristic that characterizes our data as falling into one of several categories, like being a red brick house, a tan brick house, or a slate house, is a classic example of how a single characteristic can map to numerous features. In this particular instance, we might describe this single characteristic with three features.
Feature engineering is the process of turning what we loosely consider to be the qualities of our observations into features. I won’t go into great detail about this procedure because we’ll be dealing with a situation where each observation has 13 characteristics, and we just use one number feature to represent each characteristic. As I mentioned before, the ultimate objective of supervised learning is to identify connections among data attributes. In actuality, we accomplish this by selecting one trait—referred to as our target—that we wish to predict from the rest. Depending on the issue you are attempting to resolve, the attribute you choose to employ as the target is entirely up to you.
If your goal is simply to describe the relationship between the number of rooms and the prices of houses, for instance, you could train a model with the number of rooms as a feature and the prices of houses as the target. In either case, the resulting model will contain a description of the relationship between these two characteristics, enabling you to say, for instance, that a house with more rooms is more expensive.
However, if your objective is to forecast the prices of homes for which no price data is available, you must set the price as your target so that, when your model has been trained, you may feed it the remaining data.
