Key takeaways
- The nearest neighbour algorithm, established in 1967, laid the groundwork for pattern recognition in machine learning.
- Machine learning algorithms enable systems to enhance performance and create mathematical models from training data.
- Survival analysis focuses on the expected duration of events and is significant in tasks like duration analysis and reliability theory.
- Evolutionary algorithms serve as population-based optimisation methods, addressing complex problems that traditional methods struggle to solve.
The 'nearest neighbour algorithm' in 1967, one of the first machine learning algorithms ever, was conceived, laying the foundations for basic pattern recognition. Today, in 2022, we are no short of brilliant algorithms like linear regression, KNN, SVM, decision trees and more. Machine learning algorithms are integral to helping computer systems improve their performance and automatically build a mathematical model using training data. It is owing to these algorithms that machines can make decisions without specific human instructions all the time. Through the years of computer science, algorithms have been developed and updated, and many of them are lost in space. Analytics India Magazine will take you down this time machine of some of the most popular yet forgotten machine learning algorithms.
Survival analysis
Survival analysis, part of statistics, analyses the expected duration of time for the occurrence of any event. It can be applied for tasks such as duration analysis or proportional questions, answering what part of the population will survive in a certain time or when will the current population become extinct, and more. In engineering, this is also known as reliability theory or reliability analysis because it helps study the reliability of any system that is going to close or destroy one day.
Evolutionary algorithms
As a subset of evolutionary computation, evolutionary algorithms are generic population-based metaheuristic optimisation algorithms. They provide a heuristic-based approach to solving problems that cannot be easily solved in polynomial time. The approaches used are inspired by biological evolution and occur in four processes: initialisation, selection, genetic operators, and termination.
Hidden Markov Models
A Hidden Markov Model (HMM) is a statistical model that describes the evolution of observable events. These events are not directly observable and depend on internal factors. A varied class of probabilistic graphical models allows us to predict a sequence of unknown variables from a set of observed variables.
Manifold learning
An approach to non-linear dimensionality reduction, manifold learning algorithms believe that the dimensionality of data sets is only artificially high. The algorithm generalises linear frameworks to be sensitive to non-linear structures in the data. This is usually unsupervised since it learns the structure from the data itself.
Traditional signal processing
Signal processing algorithm analyses, modifies and synthesises sound signals, images, and scientific measurements. These are typically built on three basic functions: Add, Multiply, and Delay. The functions are combined to build up complex algorithms in discrete-time systems. Traditional signal processing techniques include filtering, detection, estimation and frequency domain analysis.
Adaptive Resonance Theory (ART)
Adaptive resonance theory has been inspired by how the human brain processes information. The algorithm conceptualises supervised and unsupervised learning methods to train neural networks addressing pattern recognition, object identification and prediction.
Self-Organising Maps (SOMs)
A self-organising map is an unsupervised neural network trained using unsupervised learning techniques to produce a low dimensional, a discretised representation based on the input space of the training samples, known as a map. It is a method to reduce data dimensions. They apply competitive learning techniques to preserve all the topological properties within the input space.
Bayesian nonparametric literature
Nonparametric statistics are usually distribution-free or have a specified distribution. The distribution's parameters are unspecified and include descriptive statistics and statistical inference.
Edge detection
The edge detection algorithm uses mathematical methods to identify edges or curves in a digital image where the image brightness has discontinuities. This algorithm is fundamental to image processing, machine vision and feature detection and extraction in computer vision.
Hierarchical mixed-effects regression
Hierarchical modelling algorithms deal with data and their observations in a certain group. Hierarchical linear models or mixed-effect models are chosen approaches that allow researchers to account for ecological, contextual, and individual-level or compositional variables.
Gradient maps
Gradient maps provide basic visualisations of spatial areas and their associated values depending on the thematic issue. Gradient maps follow the location map style, illustrating values like the region to visualise data with spatial context.
Symbolic Regression
A type of regression analysis, symbolic regression algorithm, finds the appropriate model based on the given dataset and mathematical expressions. These initial expressions are outputs of random combinations of mathematical concepts like mathematical operators, analytic functions, constants, and state variables.
Conditional Random Fields
A class of statistical modelling, conditional random fields algorithms help in pattern recognition and structured prediction. CRF models the predictions as graphical models to represent the various dependencies between predictions and account for different variables.
Polynomial regression
A polynomial regression algorithm is a type of regression analysis. The regression algorithm models the relationship between a dependent (y) and independent variable (x) as 'nth' degree polynomial. Essentially, it provides the best approximation of the relationship between the dependent and independent variables.
Multi-dimensional scaling
The multi-dimensional scaling algorithm is used to visualise the similarity level of individual cases of a dataset. This algorithm falls under the branch of unsupervised ML and is a good technique to preserve the global and local structures of high-dimensional data.
