Machine learning is a tool used in analytics for using data to make decision for action. This field of study is at the crossroads of regular academic research and action research used in professional settings. This juxtaposition of skills has led to exciting new opportunities in the domains of academics and industry.
This post will provide information on basic types of machine learning which includes predictive models, supervised learning, descriptive models, and unsupervised learning.
Predictive Models and Supervised Learning
Predictive models do as their name implies. Predictive models predict one value based on other values. For example, a model might predict who is mostly likely to buy a plane ticket or purchase a specific book.
Predictive models are not limited to the future. They can also be used to predict something that has already happen but we are not sure when. For example, data can be collect from expectant mothers to determine the date that they conceived. Such information would be useful in preparing for birth .
Predictive models are intimately connected with supervised learning. Supervised learning is a form of machine learning in which the predictive model is given clear direction as to what it they need to learn and how to do it.
For example, if we want to predict who will be accept or rejected for a home loan we would provide clear instructions to our model. We might include such features as salary, gender, credit score, etc. These features would be used to predict whether an individual person should be accepted or reject for the home loan. The supervisors in this example or the features (salary, gender, credit score) used to predict the target feature (home loan).
The target feature can either be a classification or a numeric prediction. A classification target feature is a nominal variable such as gender, race, type of car, etc. A classification feature has a limited number of choices or classes that the feature can take. In addition, the classes are mutually exclusive. At least in machine learning, someone can only be classified as male or female, current algorithms cannot place a person in both classes.
A numeric prediction predicts a number that has an infinite number of possibilities. Examples include height, weight, and salary.
Descriptive Models and Unsupervised Learning
Descriptive models summarizes data to provide interesting insights. There is no target feature that you are trying to predict. Since there is no specific goal or target to predict there are no supervisors or specific features that are used to predict the target feature. Instead, descriptive models use a process of unsupervised learning. There are no instructions given to model as to what to do per say.
Descriptive models are very useful for discovering patterns. For example, one descriptive model analysis found a relationship between beer purchases and diaper purchases. It was later found that when men went to the store they often would be beer for themselves and diapers for their small children. Stores used this information and they placed beer and diapers next to each in the stores. This led to an increase in profits as men could now find beer and diapers together. This kind of relationship can only be found through machine learning techniques.
The model you used depends on what you want to know. Prediction is for, as you can guess, predicting. With this model you are not as concern about relationships as you are about understanding what affects specifically the target feature. If you want to explore relationships then descriptive models can be of use. Machine learning models are tools that are appropriate for different situations.