April 21, 2020
Depending on who you ask you will get somewhat different definitions of what machine learning is. That should not be too surprising given that it is a hot and rapidly developing field. In addition, there is always a temptation to label something machine learning to elevate the status of a project regardless of whether it qualifies as machine learning or not. The tendency to use the name machine learning as a marketing label for a project or service contributes to the fuzziness of what constitutes true machine learning.
One common misconception is to define machine learning by the algorithm or technique used such as Random Forest or XGBoost being machine learning while logistic regression or other classical statistical techniques are not. In reality, almost any modeling technique can qualify as machine learning depending on how it is being used although some techniques might be easier and more powerful than others in terms of implementing actual machine learning.
Commonly used definitions for Machine learning can be placed in two categories. The first one focuses on the ability to fit a model without specific instructions, the definition below from SAS would fit into that category
“Machine learning is a method of data analysis that automates analytical model building.”
Under this definition one could make the argument that Random forest is machine learning while logistic regression is not. However, if we combine logistic regression with code to automate the variable selection, data imputation and variable transformation process, it would fit the above definition.
The second category of definitions focuses on the ability to learn and improve as new data become available in cases where the model is not static but change over time with the presence of new information. The definition below succinctly summarizes this concept.
“Machine learning is the concept that a computer program can learn and adapt to new data without human interference.”
Although this definition is in a way stricter than the former, a simple regression model with its parameters automatically tuned to adapt to new data would qualify as machine learning. However, a Random Forest model deployed without a mechanism to automatically update with new data would not qualify as machine learning.
Under either definition, machine learning is a type of data based predictive models attempting to predict future outcomes and can be used to solve a wide range of business problem. How does this relate to Artificial Intelligence (AI)?
Most definitions for AI are similar to the one provided by Britannica as shown below:
Artificial intelligence (AI), the ability of a digital computer or computer-controlled robot to perform tasks commonly associated with intelligent beings.
For something to be considered Artificial Intelligence, it requires certain level of intelligence to perform an action which can be informed by some form of machine learning model. In this context I would argue that the second definition of machine learning is the more appropriate one. Without the ability to improve the model based on outcomes and new data, I would not consider the system particularly intelligent, no matter what algorithm was used to fit the data.
Here is a real marketing example to fit it all together. A direct mail targeting model developed on historical campaign data and used to score a lead file would not really be considered machine learning regardless of technique used, at least not under the second more strict definition.
On the other hand, an online targeting model that continuously updates as consumers access the website and buy products would be machine learning. If algorithm is developed to use this model to automatically place bids on different search words and optimize bidding relative to a given marketing budget, then this is an AI system.