Machine learning: When to use each method and technique

What exactly can you do with machine learning? We explain the various methods and techniques available to you

Machine learning: When to use each method and technique
Thinkstock

You’re probably hearing more and more about machine learning, a subset of artificial intelligence. But what exactly can you do with machine learning?

The technology encompasses a number of methods and techniques, and each has a set of potential use cases. Enterprises would do well to examine them before plunging ahead with plans to invest in machine learning tools and infrastructure.

Machine learning methods

Supervised learning

Supervised learning is ideal if you know what you want a machine to learn. You can expose it to an enormous set of training data, examine the output, and tweak parameters until you get the results you expect. Later, you can see what the machine has learned by having it predict the results for a set of validation data it hasn’t seen before.

The most common supervised learning tasks involve classification and prediction, or regression.

Supervised learning methods can be used for applications such as determining the financial risk of individuals and organizations, based on past information about financial performance. They can also provide a good sense of how customers will act or what their preferences are based on previous behavior patterns.

For example, online loan marketplace Lending Tree is using an automated machine learning platform from DataRobot to customize experiences for its customers and to predict their intent based on what they’ve done in the past, says Akshay Tandon, vice president and head of strategy and analytics.

By predicting customer intent—primarily via lead scoring—Lending Tree can sort out the people who are just shopping around for a rate versus those who are actually looking for a loan and ready to apply for one. Using supervised learning techniques, it built out a classification model to define the probability of a lead closing.

Unsupervised learning

Unsupervised learning lets a machine explore a data set and identify hidden patterns linking different variables. This method can be used to group the data into clusters based on their statistical properties alone.

A good application of unsupervised learning is the clustering algorithm used to perform probabilistic record linking, a technique that extracts connections among data elements and builds on them to identify individuals and organizations and their connections in the physical or virtual world.

This is especially useful for enterprises that need to, for example, integrate data from disparate sources and/or across different business units to build a consistent and comprehensive view of their customers, says Flavio Villanustre, vice president of technology at LexisNexis Risk Solutions, a company that uses analytics to help customers predict and manage risk.

Unsupervised learning can be used for sentiment analysis, which identifies the emotional state of individuals based on their social media posts, emails, or other written feedback, says Sally Epstein, a specialist machine learning engineer at consulting firm Cambridge Consultants. The firm has seen an increasing number of companies in financial services use unsupervised learning to derive insight into customer satisfaction.

Semisupervised learning

Semisupervised learning is a hybrid of supervised and unsupervised learning. By labeling a small portion of the data, a trainer can give the machine clues as to how it should cluster the rest of the data set.

Semisupervised learning can be used to detect identity fraud, among other uses. Fortunately, fraud is not as frequent as nonfraudulent activity, Villanustre notes, and as such fraudulent activity can be considered an “anomaly” in the universe of legitimate activity. Still, fraud exists and semisupervised anomaly-detection machine learning methods can be used to model solutions to these types of problems. This type of learning is deployed to identify fraud in online transactions.

Semisupervised learning can also be used when there’s a mixture of labeled and unlabelled data, which is often seen in large enterprise settings, Epstein says. Amazon has been able to enhance the natural language understanding of its Alexa offering by training AI algorithms on a mix of labeled and unlabelled data, she says. This has helped increase the accuracy of Alexa’s responses, she says.

Reinforcement learning

With reinforcement learning, you let the machine interact with its environment (for example, pushing damaged products off a conveyor into a bin) and provide a reward when it does what you want. By automating the calculation of the reward, you can leave the machine to learn in its own time.

One use case for reinforcement learning is the sorting of clothing and other items at a retail establishment.

Some apparel retailers have been piloting new types of technology such as robotics to help sort items such as clothing, shoes, and accessories, says David Schatsky, an analyst at consulting firm Deloitte who focuses on emerging technology and business trends.

The robots use reinforcement learning (as well as deep learning) to figure out how much pressure they should use when grasping items and how best to grab these items in inventory, Schatsky says.

A variation of reinforcement learning is deep reinforcement learning, which is well-suited for autonomous decision-making where supervised learning or unsupervised learning techniques alone can’t do the job.

Deep learning

Deep learning performs learning types such as unsupervised or reinforcement learning. Broadly speaking, deep learning mimics some aspects of how people learn, mainly by using neural networks to identify characteristics of the data set in more and more detail.

Deep learning, in the form of deep neural networks (DNN), has been used to accelerate high content screening for drug discovery, Schatsky says. It involves applying DNN acceleration techniques to process multiple images in significantly less time, while extracting greater insight from image features that the model ultimately learns.

This machine learning method is also letting many companies fight fraud, improving detection rates by using automation to detect wrongdoing.

Deep learning can also be used in the auto industry. One company has developed a neural network-based system that allows early detection of problems with cars, Schatsky says. This system can recognize noises and vibrations, and it uses any deviations from the norm to interpret the nature of the failure. It can become part of predictive maintenance, because it determines the vibrations of any moving parts of the car and can notice even minor changes in their performance.

Machine learning techniques

Neural networks

Neural networks are designed to mimic the structure of neurons in human brains, with each artificial neuron connecting to other neurons inside the system. Neural networks are arranged in layers, with neurons in one layer passing data to multiple neurons in the next layer, and so on. Eventually they reach the output layer, where the network presents its best guesses to solve a problem, identify an object, and so on.

Use cases for neural networks cross a range of industries:

  • In life sciences and health care, they can be used to analyze medical images to speed up diagnostic processes and for drug discovery, Schatsky says.
  • In telecom and media, neural networks can be used for language translations, fraud detection, and virtual assistant services.
  • In financial services, they can be used to fraud detection, portfolio management, and risk analysis.
  • In retail, they can be used to eliminate checkout lines and personalize customer experience.

Decision trees

A decision tree algorithm aims to classify items by identifying questions about their attributes that will help decide in which class to place them. Each node in the tree is a question, with branches leading to more questions about the items, with the leaves being the final classifications.

Use cases for decision trees can include building knowledge management platforms for customer service, pricing predictions, and product planning.

An insurance company might use a decision tree when it requires insights into what type of insurance products and premium adjustments are needed based on potential risk, says Ray Johnson, chief data scientist at business and technology consulting firm SPR. Using location data overlaid with weather-related loss data, it can create risk categories based on submitted claims and expenditure amounts. Then it can evaluate new applications for coverage against models to provide a risk category and the potential financial impact, he says.

Random forests

While a single decision tree must be trained to provide accurate results, the random forest algorithm takes an ensemble of randomly created decision trees that base their decisions on different sets of attributes, and lets them vote on the most popular class. 

Random forests are versatile tools for finding relationships in data sets and are quick to train, Epstein says. For example, unsolicited bulk email has long been a problem, not just for users but also for the internet service providers that have to manage the increased load to servers. As a response to this problem, automated methods for filtering spam from normal email have been developed, using random forests to quickly and accurately identify unwanted email, she says.

Other uses for random forests include identifying a disease by analyzing a patient’s medical records, detecting fraud in banking, predicting call volume in call centers, and forecasting profits or losses through the purchasing of a particular stock.

Clustering

Clustering algorithmsuse techniques such as K-means, mean-shift, or expectation-maximization to group data points based on shared or similar characteristics. This is an unsupervised learning technique that can be applied to classification problems.

The clustering technique is particularly useful when needing to segment or categorize, Schatsky says. Examples include segmenting customers by distinct characteristics to better assign marketing campaigns, recommending news articles to certain readers, and effective police enforcement.

Clustering is also effective for discovering groupings in complex data sets that may not be obvious with the human eye. Examples range from categorizing similar documents in a database to identifying crime hot spots from crime reports, Epstein says.

Association rule learning

Association rule learning is an unsupervised technique used in recommendation engines, which looks for relationships between variables.

This is the technique behind the “people who bought X also bought Y” suggestions on many e-commerce sites, and examples of how this is being used are common.

A specific use case might be a specialty food retailer that wants to drive additional sales, Johnson says. It would use this technique to examine customer buying behavior to provide special tins and bundles for products celebrating events, sports teams, and so on. The association rules techniqueprovidesinsightsthat can uncoverwhen and where customers bought the preferred combination of products.

Using information on past purchases and timeframes letsthe companyproactively create a rewards program, Johnson says, and provide special customized offersto drive future sales.

Copyright © 2018 IDG Communications, Inc.

How to choose a low-code development platform