Basic types of machine learning
Broadly speaking, machine learning is a sort of sub-field of artificial intelligence where we get a computer to learn things on their own. A computer has a learning algorithm which helps it identify certain patterns in the data it is presented. This helps it “predict” things and build models without being pre-programmed to do so with any specific rule. Some companies that provide software infrastructure can help you with your machine learning use cases to cater to your specific needs.
However, in order to start implementing machine learning, it is good to explore types of it and find out which is most suitable for your business. In this article, we will explore the three major types of machine learning.
Unsupervised learning
This approach allows us to deal with problems with little to no idea what our outcomes will be. Using this type of learning we derive a structure from the data where the effect of the variables are unknown. The data in this type of learning doesn’t require it to be labelled. The algorithm sifts through the unlabeled data to look for any patterns it can use to group data points into subsets. When we think of deep learning (including neural networks) we think of this type of learning.
A more concrete example of unsupervised learning includes grouping customers by their purchasing behaviour. Classifying individuals by their different interests and identifying associations in customer data. For example, when someone buys a certain type of jacket they may also be interested in a certain type of glasses or shoes. It is quite an interesting fact that only 4% of businesses make an effective use of the data at their disposal.
Supervised learning
In stark contrast with unsupervised learning, in supervised learning, we are given a specific data set with our desired output already in mind. This is one of the main reasons why supervised learning is so effective in fraud detection, inventory optimisation and sales forecasting. This type of learning is good for binary classification, multi-class classification, regression modelling and ensembling.
For example, a model could be fed data from countless bank transactions some labelled fraudulent and some not. The algorithm can then pick up on patterns and in time learn when a transaction raises certain flags. This can also work in the healthcare industry when ML (machine learning) can identify risk factors for diseases and plan preventing measures accordingly. Predict whether or not people will vote for a given candidate, housing prices, loan risk, etc.
Reinforcement learning
Reinforcement learning works on a reward-punishment basis. Instead of one input providing one output, the learning algorithm produces multiple ones and is set to choose the one set with certain variables. Punishments are identified as variables that are outside its chosen goal so the algorithm avoids them. Reinforcements learning is used in a lot of areas like Robotics, Video game AI, resource management, CRM software, Self Driving cars, etc.
The best example we can provide is the famous (or maybe now infamous) Facebook’s News Feed. Facebook uses reinforced learning so it can personalize every person’s news feed to cater to their individual likes. The algorithm can detect what kind of content the user interacts with most and places similar ones in their feed. This way the algorithm is reinforcing the user’s patterns in online behaviour.
Conclusion
Our reliance on tools is what makes us who we are today. Machine learning is just the latest tool in our already big toolbox, it has its upsides and downsides. The upside is that using machine learning companies can understand their customers on a level never before imaginable. By collecting patterns of behaviour and correlating it with customer data, machine learning algorithms can tailor their products or services to better fit their customer’s needs and demands.
Google users machine learning to better place their advertisements, Uber uses it to match drivers with riders. But there are downsides to this as well. First of all machine learning is expensive, as it’s usually driven by data scientists with high salaries. Second, these projects also require specific software infrastructure, which can also be high cost. The darkest yet, it may happen that certain population exclusions or errors in data sets occur, and can lead to all sorts of discriminations or inaccurate world models. Like with any tool we must embrace the good side but also be mindful of the bad.