Machine Learning Algorithms Categories

Machine Learning Algorithms learn from data as humans learn from experience.

But the type of learning and the goal varies from algorithm to algorithm. Some algorithms need prior knowledge of the target data based on all the input data.

When a lot of target data and predictor data is fed, the machine learns from this and then predicts the target output for new data. Some other algorithms do not need any target data. It just gleans insights from existing data based on common patterns within the data. They turn out pretty powerful in bringing forward unseen insights to the business and show a lot of opportunities for innovation and market growth. There are yet others that can learn better based on rewards earned.

Here is one way of categorising ML Algorithms and a high level overview of each of them.

Regression

Regression algorithms are one of the most basic algorithms that find immense usage across various fields. It models the relationship between a dependent (target) variable and one or more independent (predictor) variable(s). This is all about finding a best fit line between predictors and target variables

Here the target variable is a continuous variable like age, price, temperature etc. which are to be predicted.

An example: Given the amount spent on TV advertisement, you can predict the sales increase or decrease as depicted in the graph here

An example algo: Multi-Linear Regression

#SupervisedAlgorithm

Classification

This is again one of the fundamental types of algorithms used for predicting discrete target variables like diabetic or not, will leave the company or not, spam or ham etc. It is a task of assigning classes or labels to a given data point.

Here the output is a categorical variable. The number of categories that you want to classify can be more than two as well, termed as Multi-class Classification as against Binary Classification.

An Example Algo: Logistic Regression

#SupervisedAlgorithm

Clustering

Clustering Algorithms are a set of algorithms that help in discovering interesting patterns in data, such as customers with similar tastes, similar behaviours. It looks for natural groups based on the given input data, thus the name clustering.

Here, there is no predefined notion of a label given to an input. One powerful application could be Clustering of News which can show real news from fake news.

Some other applications are Customer Segmentation leading to cross-sell and up-sell success, detecting Fraudulent activity or even categorising network traffic.

A typical clustering plot using scatter plot is shown here. This could be two groups of customers based on some shopping patterns, giving you an insight into whom to invest more.

An Example Algo: K-Means Algorithm

#UnsupervisedAlgorithm

Tree Models

Tree Models can be used for Classification or Regression but work on a series of if-then-else rules that help in the predictions. This resembles an up-side-down tree that is used to make decisions.

'if-then-else' rules are not new to the Computer world. However, so far, humans had these rules in mind based on domain expertise and gave that as a 'Business Requirement' to software developers to code the same. However, Tree Models are able to intelligently discover these rules, by just looking at given data. This makes the rules dynamic with changing data being fed to models.

A decision tree splits data into multiple data sets and uses each subset of data to arrive at a decision. The creation of the best subsets of data and the rules that help in final conclusions are all machine learnt.

For example, given a list of parameters about a patient, the model predicts if he/she is a diabetic or not, based on a set of rules, it has auto generated "seeing" past data, as shown in the figure below: i.e it shows that if some one is above 54.5 years, is a male and has cholesterol equal to or less than 272.5, they are going to be diabetic.

An example Algo: Decision Trees

#SupervisedAlgorithm

Time Series Models

Time Series models deal with data that have a time element strongly associated with it. This is a method where past data is used to predict the future. It takes into consideration, things like seasonality, repeated patters over time periods like weeks, quarters or months and helps in forecasting.

Examples include forecasting the retail sales in the future based on sales behaviour over a historical period of significance. Similarly, forecasting flight bookings.

An example Algo: ARIMA (Auto Regressive Integrated Moving Average)

NOTE: Forecasting Vs Prediction

Forecasting and prediction are not synonymous in the ML world. They have a technical difference. 'Forecasting' is about extrapolation beyond the period for which the data is available while 'prediction' is about interpolation of the target data within the range of data available.

Deep Learning

This is the science of using Neural Networks to perform very complex tasks. This is inspired by how the human brain works and is used for tasks such a speech recognition, image classification and machine translation.

These models consist of cells or neurons that receive a signal and based on a specific threshold either amplify and pass the signal to the next neuron or inhibit that signal as required by the output - just like the human brain.

This is quite advanced in Machine Learning and is a deep branch by itself.

An example Algo: Convolutional Neural Networks

Conclusion

Hope this gives you a basic understanding of the broad categories of machine learning models that exist. This is clearly not an exhaustive list but some of the most often used categories.