To avoid repetition about what is Machine Learning and why it matters, and if it is our first article that you are reading, here you can find our last post covering these points.
Now to go straight to the subject, let’s divide Machine Learning into three main categories: Supervised Learning, Unsupervised Learning, and Reinforcement Learning. The difference between the three categories, in the same order, is like: a kid being helped by parents to take every decision, a kid taking a decision based on his understanding, and a kid taking a decision in a new situation based on past experiences.
1. Supervised Learning – It is about teaching the machine by example: providing the algorithm with inputs and outputs, for it to make predictions from identified patterns and observations, and then correcting these predictions until a high level of accuracy is achieved. This can be achieved using:
- Classification: The algorithm determines what category the observations belong to, after drawing a conclusion from the observed values. Will it be cold outside tomorrow? Will we have rainfall?
- Regression: The algorithm focuses on one dependent variable, and the relationship between it and other variables to predict an outcome in form of real value. While Classification was about whether it will be cold tomorrow or it will rain, here it is more about what the temperature be like tomorrow and what will be the value of rainfall in cm.
Linear Regression, Logistic Regression, SVM (Support Vector Machine), Decision Tree, Naive Bayes, and KNN (K Nearest Neighbors) algorithms are related to Supervised Learning.
2. Unsupervised Learning – Unlike Supervised Learning, there is no human interaction here. The machine itself groups and organizes large sets of data, for it to identifies patterns and determine correlations and relationships. As it assesses the data, the ability of the machine to make decisions gradually improves. This can be done with:
- Association: Discovering the probability of co-occurrence of items in a collection. If a customer buys coffee and milk, he is most likely to buy sugar.
- Clustering: Grouping similar data based on defined criteria together, then analyzing each group to find patterns. This can be the case in medical researches and Targeted Marketing.
K-Means, Apriori, and PCA (Principal Component Analysis) algorithms are related to Unsupervised Learning.
3. Reinforcement Learning – Providing the machine with different parameters for it to try different options and possibilities, and determine the most optimal result. It is about learning the machine trial and error, so it learns from past experiences and begins to adapt to achieve the best result possible. This kind of Machine Learning is usually used in robotics where a robot bumping into obstacles receives negative feedback (error).
Upper Confident Bounce (UCB) and Thompson Sampling are related to Reinforcement Learning.
So, which is the best algorithm to choose?
Accuracy, training time, ease of use, linearity, number of parameters, number of features, special cases, etc. are some aspects that should be taken into consideration when selecting the algorithm(s) to work with.
Choosing the algorithm that will work best is hard to define at the start, so it is better to run the data on different algorithms and select the one that performed the best. However, while someone can consider accuracy first, his colleague may prefer to work on algorithms he knows best.
Computer Engineer • Entrepreneur • Blogger