One of the most interesting fields of quantitative research today is machine learning. Machine Learning serves the purpose of developing models through the use of algorithms in order to inform action.

Examples of machine learning are all around us. Machine learning is used to make recommendations about what you should purchase at most retail websites. Machine learning is also used to filter spam from your email inbox. Each of these two examples involves making a prediction about previous behavior. This is what machines learning does. It learns what will happen based on what has happened. However, the question remains how does a machine actually learn.

The purpose of this post is to explain how machines actually “learn.” The process of learning involves three steps which are…

1. Data input
2. Abstraction
3. Generalization

Data Input

Data input is simply having access to some form of data whether it be numbers, text, pictures or something else. The data is the factual information that is used to develop insights for future action.

The majority of a person’s time in conducting machine learning is involved in organizing and cleaning data. This process is actually called data mugging by many. Once the data is ready for analysis, it is time for the abstraction process to begin.

Abstraction

Clean beautiful data still does not provide any useful information yet. Abstraction is the process of deriving meaning from the data. The raw data represent knowledge but as we already are aware the problem is how it is currently represented. Numbers and text mean little to us.

Abstraction takes all of the numbers and or text and develops a model. What a model does is summarize data and provides an explanatiofout the data. For example, if you are doing a study for a bank who wants to know who will default on loans. You might discover from the data that a person is more likely to default if they are a student with a credit card balance over $2,000. How this information is shared with the researcher can be in one of the following forms

• equation
• diagram
• logic rules

This kind of information is hard to find manually and is normally found through the use of an algorithm. Abstraction involves the use of some sort of algorithm. An algorithm is a systemic step-by-step procedure for solving a problem. It is very technical to try and understand algorithms unless you have a graduate degree in statistics.   The point to remember is that algorithms are what the computer uses to learn and develop models.

Developing a model involves training. Training is achieved when the abstraction process is complete. The completion of this process depends on the criteria of what a “good” model is. This varies depending on the requirements of the model and the preference of the researcher.

Generalization

A machine has not actually learned anything until the model it developed is assessed for bias. Bias is a result of the educated guesses (heuristics) that an algorithm makes to develop a model that is systematically wrong.

For example, let’s say an algorithm learns to identify a man by the length of their hair. If a man has really long hair or if a woman has really short hair the algorithm will misclassify them because each person does not fit the educated guess the algorithm developed. The challenge is that these guesses work most of the time but they struggle with exceptions.

The main way of dealing with this is to develop a model on one set of data and test it on another set of data. This will inform the researcher as to what changes are needed for the model.

Another problem is noise. Noise is caused by measurement error data reporting issues trying to make your model deal with noise can lead to overfitting which means that your model only works for your data and cannot be applied to other data sets. This can also be addressed by testing your model on other data sets.

Conclusion

A machine learns through the use of an algorithm to explain relationships in a data set in a specific manner. This process involves the three steps of data input, abstraction and generalization. The results of a machine learning model is a model that can be use to make prediction about the future with a certain degree of accuracy.