eWEEK content and product recommendations are editorially independent. We may make money when you click on links to our partners. Learn More.
If you use the Internet in any capacity, you will inevitably run into algorithms. From Google’s search engine to Facebook’s timeline algorithms to the systems that help financial institutions process transactions, algorithms are the foundation of artificial intelligence.
Despite being core to our digital lives, algorithms aren’t often understood by anyone besides the people who create them. Infamously, despite supporting nearly 400,000 full-time creators with its platform, Youtube’s algorithm – it recommends videos and spotlights channels related to users’ interests – is known for being an oblique black box by which creators feast and famine.
This article will shine a light on this fundamental aspect of the tech industry.
Also see: Top AI Software
What is an Algorithm?
In basic terms, an algorithm is a set of solidly-defined steps which need to be taken in order to reach a planned result. In particular, it is used to solve mathematical equations. It can be broken up into three broad components:
- Input: The information you already know at the beginning of the problem.
- Algorithm: The sequence that needs to be followed step-by-step to achieve.
- Output: The expected results if all steps in the sequence are followed to the letter.
An example of an algorithm-like system outside of the tech world would be cooking recipes. You have your input (the ingredients), you have your algorithm (the steps of the recipe which need to be followed more or less exactly), and you have your output (a hopefully-edible dish).
We’re not kidding when we say algorithms are part of the atomic structure of our digital lives, either. Any computer program you utilize is running multiple algorithms to perform its functions. From your web browser to your word processor to the Microsoft Solitaire that has been included with Windows since 3.0, every single one of them runs off of algorithms.
Also see: The Future of Artificial Intelligence
How Do Algorithms Work in AI?
Fundamentally, artificial intelligence (AI) is a computer program. Meaning that, like Firefox or Microsoft Word or Zoom or Slack, any AI or machine learning (ML) solution you come across will be built from the ground-up with algorithms.
What algorithms do in AI, as well as machine learning, is variable. Broadly speaking, they define the rules, conditions, and methodology an AI will use when processing and analyzing data. This can be as simple as defining the steps an AI needs to take to process a single invoice to having an AI filter out pictures with dogs among a dataset containing hundreds of thousands of pictures.
Algorithms in machine learning help predict outputs even if given unknown inputs. AI algorithms function similarly by solving different categories of problems. The types of problems that AI algorithms solve can be divided into three broad categories:
- Classification: A type of machine learning which is used to predict what category, or class, an item belongs to. One example would be programming an AI to differentiate between spam messages and messages you actually need.
- Regression: A type of machine learning which is used to predict a digital label based on how an object functions. One example would be using historical data to forecast stock market prices and projections.
- Clustering: A type of machine learning which is used to sort objects into groups based on similarities in their functionality. One example would be using an algorithm to sort through a set of financial transactions and picking out instances of potential fraud.
Also see: How AI is Altering Software Development with AI-Augmentation
Types of AI Algorithms
Classification Algorithms
Below are some examples of classification algorithms used in AI and machine learning.
Binary Logistic Regression
Binary logistic regression can predict a binary outcome, such as Yes/No or Pass/Fail. Other forms of logistic regression, such as multinomial regression, can predict three or more possible outcomes. Logic regression can often be found in use cases like disease prediction, fraud detection, and churn prediction, where its datasets can be leveraged to assess risks.
Naive Bayes
Naive Bayes is a probability algorithm built off of incorporating independence assumptions into its models, meaning it operates off the assumption that no two measurements in a dataset are related to each other or affect each other in any way. This is why they’re called “naive.” It’s commonly used in text analysis and classification models, where it can sort words and phrases into specified categories.
K-nearest Neighbors (k-NN)
While also sometimes used to solve regression problems, k-NN is most often used to solve classification problems. When solving classification problems, it separates data points into multiple classes onto a plane to predict the class label of a new data point. The new data point is given a new classification based on which class label is most often represented around it on the plane. k-NN is also known as a “lazy learning” algorithm, which means it doesn’t undergo a full training step, instead only saving a training dataset.
Decision Tree
A supervised learning algorithm, decision trees can also be used for either classification problems or regression problems. It’s called a “tree” because it possesses a hierarchical structure. Starting with a root node, it branches out into smaller internal or decision nodes where evaluations are conducted to produce subsets, which are represented by terminal or leaf nodes.
An example would be starting with a root node for martial arts which are then split into internal nodes for martial arts with a striking focus and martial arts with a grappling focus. These internal nodes can then split into terminal nodes for specific martial arts like boxing, jiu-jitsu, and Muay Thai. These algorithms are great for data mining and knowledge discovery tasks because they’re easy to interpret and require very little data preparation to be deployed.
Random Forest
Random forests leverage the output of multiple decision trees to produce a prediction. Like decision trees, random forests can be used to solve both classification and regression problems. Each tree is made up of a data sample drawn from a training dataset that uses sampling with replacement. This adds randomization to the decision trees, even if they draw from the exact same dataset.
In classification problems, a majority vote is determined from the output of these randomized decision trees. For example, say there are 10 decision trees dedicated to determining what color a dress is. Three sets say it is blue, two sets say it is black, four sets say it is pink, and one set says it is red. The dress would be categorized as pink.
Random forests are the algorithm of choice for finance-focused machine learning models, as it can lower the time taken for pre-processing and data management tasks. Fraud detection, option pricing, and customer credit risk evaluation are all examples of its use in finance. The random forest algorithm is trademarked by Leo Breiman and Adele Cutler.
Also see: Best Machine Learning Platforms
Regression Algorithms
Below are some examples of regression algorithms used in AI and machine learning.
Linear Regression
An algorithm with use in both statistics and the social sciences, linear regression is used to define the linear relationship between a dependent variable and an independent variable.The goal of this sort of algorithm is to determine a possible trend line with the given data points. Businesses often use linear regression when determining how revenue is affected by advertising spending.
Poisson Regression
Poisson regression is a type of regression where a predicted variable is always assumed to follow a Poisson distribution. A Poisson distribution is a probability function that can help determine the probability of a given number of events happening within a specific, fixed time period.
For example, you could use Poisson regression to determine how likely a classroom of high schoolers is to solve a Rubik’s Cube within 24 hours. Or, you could predict how likely a restaurant is to have more customers on specific days based on the average number of diners they serve in a week.
Ordinary Least Squares (OLS) Regression
One of the most popular regression algorithms, OLS regression takes ordinal values as input to determine the linear relationship between multiple variables. The algorithm is most useful when predicting the likelihood of something being ranked on an arbitrary scale, such as how likely a game is to be rated a 7 on a scale of 1–10. It’s often used in the social sciences, since surveys in that field frequently ask participants to evaluate something on a scale. OLS regression is also known as ranking learning.
Lasso (Least Absolute Selection and Shrinkage Operator) Regression
Lasso regression takes an OLS regression and adds a penalty term to the equation. This can help you create a more complex representation of data than is otherwise possible with simple OLS. It can also make the representation more accurate. Lasso regression is also known as L1 regularization.
Neural Network Regression
Neural networks are one of the most popular methods of AI and ML training out there. As the name implies, they’re inspired by the human brain and are great at handling datasets that are too large for more common machine learning approaches to consistently handle.
Neural networks are a versatile tool and can perform regression analysis as long as they are given the appropriate amount of prior data to predict future events. For example, you could feed the neural network customers’ web activity data and metadata to determine how likely a customer is to leave your website without buying anything.
Check Out: Top Predictive Analytics Solutions
Clustering Algorithms
Below are some examples of clustering algorithms used in AI and machine learning.
K-Means Clustering
An unsupervised learning algorithm, k-means clustering takes datasets with certain features and values related to these features and groups data points into a number of clusters. The “K” stands for the number of clusters you’re trying to classify data points into. K-means clustering possesses a number of viable use cases, including document classification, insurance fraud detection, and call detail record analysis.
Mean Shift Clustering
A simple, flexible clustering technique, mean shift clustering assigns data points into clusters by shifting points toward the area with the highest density of data points (called a mode). How a cluster is defined in this setting can be dependent on multiple factors, such as distance, density, and distribution. It’s also known as a “mode-seeking algorithm.” Mean shift clustering has uses cases in fields like image processing, computer vision, customer segmentation, and fraud detection.
Density-Based Spatial Clustering of Applications with Noise (DBSCAN)
DBSCAN separates high-density clusters from one another at points of low data point density. Netflix’s movie recommendation algorithm uses a similar clustering method to determine what to recommend to you next.
For example, if you watched the recent Netflix movie “Do Revenge,” the algorithm would look at other users who also watched “Do Revenge” and suggest movies and shows based on what those users watched next. DBSCAN is excellent at handling outliers in datasets. Viable use cases for DBSCAN include customer segmentation, market research, and data analysis.
Balanced Iterative Reducing and Clustering using Hierarchies (BIRCH)
BIRCH is a clustering technique often used for handling large datasets. It can scan an entire database in a single pass and focuses on spaces with high data point density within the database and provides a precise summary of the data.
A common way to implement BIRCH is to do so alongside other methods of clustering that can’t handle large datasets. After BIRCH produces its summary, the other clustering method runs through the summary and clusters that. As such, BIRCH’s best use cases are for large datasets that normal clustering methods cannot efficiently process.
Gaussian Mixture Model (GMM)
Much like Poisson regression utilizes the concept of Poisson distribution, GMM models datasets as a mixture of multiple Gaussian distribution models. Gaussian distribution is also known as “normal distribution,” and as such, it can be intuitive to assume that a dataset’s clusters will fall along the lines of a Gaussian distribution.
GMMs can be useful for handling large datasets, as it retains many of the benefits of singular Gaussian models. GMM has found use in speech recognition systems, anomaly detection, and stock price prediction.
Want to See What Exciting Things Companies Are Doing With AI Algorithms? Take a Look at Top Natural Language Processing Companies 2022