The Three Types of Machine Learning Algorithms
Machine learning and Artificial Intelligence (AI) are probably the most used buzzwords today. However, most of the journalists, consumers and businesses don’t really understand the two terms sufficiently well. There are several definitions of AI, with the minimum definition requiring the agent to “see”, “hear” or “read” and take singular decisions based on it. The most complex definition of AI requires the agent to replicate general human intelligence as seen in movies. However, for most business contexts, the minimum intelligence already provides far higher efficiency than what exists in the current processes. Most of this can be achieved through the use of Machine Learning.
Machine Learning is a domain of computer science with its base in computational mathematics and statistics. The machine is shown a ton of data and it learns the pattern in the data to make future predictions, recognise new patterns or suggest different classes to the data. Broadly speaking, Machine Learning algorithms are of three types- Supervised Learning, Unsupervised Learning, and Reinforcement Learning. A basic understanding of the different types of algorithms will help you choose an algorithm for your project or will simply help you appreciate the vast variety of problems AI can solve using Machine Learning.
When the machine is supervised while it is “learning”, the training type is called supervised learning. But what does supervising a machine really mean? It means that we provide the machine with a ton of information about a case and also provide it with the case outcome. The outcome is called the labelled data while the rest of the information is used as input features. For example, we show the machine 1000 cases when customers defaulted on a loan and 1000 cases where customers did not default. Here default / not-default is the outcome and hence the labelled data while all the other characteristics like age, salary, loan amount, outstanding amount, other loan history etc. are the input features. In this case, the machine is supervised by the labelled data to learn what the relationships and dependencies are between the outcome of default and the borrower information.
There are two further categories into which we can divide supervised learning: regressions and classifications, and each has its own set of use cases and merits. Common supervised learning algorithms include: Linear regression; Naïve Bayes, Nearest Neighbours, Decision Trees, Support Vector Machines and Neural Networks.
As the name suggests, in case of unsupervised learning, there is no help from the user for the computer to learn. In the lack of labelled training sets, the machine identifies patterns in the data that is not so obvious to the human eye. So, unsupervised learning is extremely useful to recognise patterns in data and help us take decisions. For example, if we didn’t know which customers defaulted on loans, and but fed the borrower information to the machine, it would be able to pick out similar patterns among the different borrowers and grouped them in 3-4 buckets or clusters. Unsupervised learning is also often used for anomaly detection, like to uncover fraudulent transactions or payments. The most common use of unsupervised learning is in clustering problems, with the most talked about algorithms being k-means and hierarchical clustering, though other algorithms like Hidden Markov models, Self-Organizing Maps or Gaussian Mixture models are also often used.
Reinforcement learning is probably the closest to how we as humans learn. In this case, the algorithm or the agent learns continually from its environment by interacting with it. It gets a positive or a negative reward based on its action. Let’s consider the same example of customers with bank loans. A Reinforcement Learning algorithm looks at the information of a customer and classifies him / her as a high-risk customer. When the customer defaults, the algorithm gets a positive reward. If the customer doesn’t default, the agent receives a negative reward. The reward in both the cases helps the agent understand the problem and the environment better, and thus helps to make better decisions on our behalf. Common algorithms include Q-Learning, Temporal Difference and Deep Adversarial Networks.
Reinforcement Learning is probably the hardest to execute yet in a business environment but has been commonly used for self-driving cars or the famous Alpha Go chess match trials.
Now that you know about the types of machine learning algorithms, we are sure you can appreciate the various groups of problems that can be solved with Machine Learning. So, whenever you want to think of a better way to fasten up a current human driven process, think machine learning, and you may be pleasantly surprised!