Predicting the future: Sounds like something from a movie, not like an actual possibility. But it is one of the many features of machine learning.
As the name suggests, machine learning depends on one activity: learning. It can be done in two different ways: with or without supervision. Many machine learning systems that acquire their data using supervised learning can predict the future. To do this, the machine finds regularities and creates prediction models.
The example of movies is a good way of explaining the differences between supervised and unsupervised learning. With supervised learning, movies are classified in different genres. A machine using unsupervised learning, on the contrary, will note that teenagers love to watch movies starring Ryan Gosling. Accordingly, it will offer young users movies with this actor. With supervised learning, the machine starts with pre-defined categories; with unsupervised learning, it defines its own categories.
The fundamental difference between supervised and unsupervised learning is the way the machine handles data. With supervised learning, the possible answers are known from the start, since humans entered all the information into the model. With unsupervised learning, the machine understands the data available to it.
To explain the process of supervised learning, let’s look at the call center of a telecommunication provider. Many customers contact the call center to complain that their WLAN does not work. The artificial intelligence can then suggest a router reboot to the agent. The machine makes this suggestion because it learned that this actions had often been successful in the past. It can only know this if the data records for previously successful actions have been marked, i.e. labeled. For this reason, it is important that call agents procure feedback on the success of the suggestions.
The model is trained using these labeled data records so that it is finally able to match an existing solution to a new data record. Once training is completed, this data can be used for example to find solutions for Internet trouble.
With unsupervised learning, on the contrary, the system uses unmarked and unlabeled data and classifies it in categories. Usually, this is done by means of segmentation where data is grouped according to common features, also known as ‘clustering’.
Many of the most successful and best-known ML applications, such as speech recognition, spam detection or damage assessment, simply classify records presented to them. Other ML applications even dare to "look into the crystal ball" by making predictions about future events. These include weather forecasts, stock price forecasts and predictive maintenance models.