What is Machine Learning?

I released another video on data science basics this week: What is Machine Learning?

Machine learning describes computer systems that are able to automatically perform tasks based on data. A machine learning system takes data as input and produces an approach or solution to a task as output, without the need for human intervention. Machine learning is closely tied to the idea of algorithms. An algorithm is a set of instructions, that when followed, performs a specific task. It is a like recipe that can be programmed into a computer to do useful things, like performing calculations or doing spell checking. With simple algorithms, there’s usually one correct answer we are looking for and the algorithm carries out the steps necessary to find it. In data science, tasks we are interested in often don’t have a single correct answer. For instance, we might want to categorize customers into groups with similar shopping habits or predict which customers are likely to buy a new product. Machine learning algorithms take data, detect patterns in it and then use those patterns to create models that help us answer these sorts of questions. This process of learning a model from data is known as model training. Once a machine learning algorithm has created a model from existing data, it can be applied to new data that it hasn’t seen yet to do things like categorizing the new observations or predicting future behaviors. Applying a model to new situations is known as model testing or prediction. Machine learning systems typically fall into one of three categories: supervised learning, unsupervised learning and reinforcement learning. 1. In supervised learning, there is a target variable that is labeled by humans for observations we’ve already seen and the goal is to predict that target for unlabeled cases that we see in the future. For instance, we might have data on past customers, including whether they purchased a product and we’d like to know whether new customers will purchase that product. In this case the target we are interested in, is whether the product is purchased and the labels are records of past purchases provided as a part of the data set used to train the model. Supervised learning problems come in two general flavors: classification, where the target variable is a category, like whether a customer is going to purchase something or not and regression, where the target is a numeric value, like how many units of a product are going to be sold. Examples of applications of supervised learning include, spam detection, stock forecasting, voice recognition and image recognition. 2. Unsupervised learning problems involve finding useful patterns in data, where there is no labeled target variable provided that the computer can learn from. Examples of unsupervised learning include clustering, which involves grouping similar observations together, anomaly detection, which involves flagging unusual observations and dimensionality reduction, which involves cutting out or combining variables to reduce the size of a data set and make it easier to work with. Applications of unsupervised learning include customer segmentation, which assigns customers to groups that exhibit similar qualities, fraud detection, which tries to identify the activity of bad actors and behavior associations, which can link actions that commonly occur together, such as purchasing milk and cereal in the same shopping trip. 3. Reinforcement Learning describes systems that act within an environment and use feedback from those actions to learn better ways of acting in the future. Applications reinforcement learning include game playing, like creating AI for video games or agents capable of defeating human players at games like chess and go, stock trading, where trading algorithms can learn from mistakes to boost returns and robotics, where robots can learn how to move more safely and efficiently within their environment. In the end, the goal of machine learning is to get computers to take in data and use it to generalize about the world: to learn from what has been observed in the past to understand new observations and use that understanding in useful ways, like making better choices.