What Questions Can Machine Learning Help You Answer?

Last updated on Oct 11, 2021

Machine learning may be able to help you and your business questions and do tasks that might be too difficult or time consuming to be done efficiently by humans. As it can be used to quickly process immense datasets, machine learning can provide data-backed answers specific to your company that could otherwise take more time and more effort for your team to come up with on their own.

The methods with which machine learning processes data are called predictive algorithms. Predictive algorithms can be either classification models or regression models. This article will review both types of predictive algorithms and go through some of the ways each of them can answer important questions for your business.

 

1. Answer "Yes or No" Questions with Classification Models

Classification models predict a binary, or discrete, outcome in which there are usually only two possible classes for a given input (though multiclass classification algorithms are also available). For example, during a spam email detection task, a classification model can be used to predict whether a given email is one of two options: either spam or legitimate. In essence, a classification model answers a simple question for a given input: Is this input Type A or Type B?

Additionally, classification models also generate a probability value that describes the likelihood that a given data point belongs to either of the potential outcome classes. To do this, the model is equipped with a default classification threshold which functions as a cutoff score that decides which class a data point is assigned to. In the email example, let’s assume the classification threshold for a spam email is 0.5. If a classification model assigns an email a 0.6 probability of being spam, it exceeds the classification threshold and will be classified as spam. If the model generates a probability of 0.4 for a given email, the model will classify the email as legitimate.

 

Real-World Applications of Classification Models

The following are some real-word applications of classification models:

  1. Medical Diagnosis

Classification models can be used to predict the probability of an individual developing a given disease. Modeling diagnoses can help busy hospitals with another tool to supplement their knowledge base when treating patients.

  1. Customer Attrition

Use a classification model to predict the probability that one of your clients will buy a competitor’s product, or the probability that they’ll stop using your product or service altogether. Common examples of attrition include a client deciding to go with a different provider or canceling a magazine subscription.

  1. Credit Card Fraud

Banks forecast the likelihood that a credit card transaction is fraudulent using classification models, saving time and money.

  1. Machine Breakdown

Classification models can be used to warn of a likely machine failure in the near future using probability and industry-specific maintenance information. With a warning ahead of failure, a technician can take action to fix the machine before it affects production.

  1. Employee Retention

Classification models can be used to calculate the likelihood that a given employee will leave your company, giving you a head start on managing your staffing needs. Predicting the potential of employees to leave can also help you improve company morale and long-term employee retention.

  1. Dating Compatibility

Ever used a dating app or service? Classification models can even predict whether or not you’ll hit it off with someone on a date.

  1. Target Marketing

Use a classification model to estimate the probability of someone buying your product or service. You can use these results to more accurately target your marketing to your customer base.

  1. Object and Speech Recognition

Classification models can be used to “recognize” objects or sounds and determine the likelihood that it was identified correctly by the model. In object and speech recognition, a model can be used to answer questions like “What animal is in this image?” or “Did the audio clip use the word ‘ball’ or ‘bawl’?”

 

2. Answer "How Much" Questions with Regression Models

In contrast to classification models, regression models can be used to predict continuous outcomes. For example, while a classification model can predict whether or not someone will buy your product, regression models can predict how much money someone will spend in your store. With regression models, a magnitude or quantity value is generated as the continuous outcome prediction.

 

Real-World Applications of Regression Models

The following is a list of real-world applications of regression models:

  1. Life Expectancy

Regression models are often used to predict how long a person is likely to live, which is useful information when setting life insurance premiums and pension annuities.

  1. Credit Loss

When someone defaults on a loan, a regression model can be used to calculate how much of the debt is likely to be written off.

  1. Spend

Regression models can be used to calculate how much money someone is likely to spend in a store within a given amount of time, which can be useful for both revenue predictions and target marketing.

  1. Call Length

Predict how long a phone call will last using regression modeling. Phone tree data can be used as training data for such a model, which could help with resource organization in a call center.

  1. Economic Forecasting

Regression models can be used to forecast how inflation or GDP will change in the future. 

  1. Occupancy

Use a regression model to generate a probability of how likely someone is to be available to take a sales call. Then, adjust the timing of your sales calls to reach customers and prospects more efficiently.

  1. Response Time

Regression models can be used to forecast how long it may take a person to reply to a given type of communication. Predict response times to emails, letters, bills, or other communications to your customers or stakeholders.

Both classification and regression models are widely used. However, classification models are typically used more frequently due to the simplicity of a classification problem. The binary outputs of classification models are generally easier for organizations to picture how a predictive model might be implemented into their current processes.

Can’t decide which model type is best for your use case? In some cases, predictive models can be combined. For example, take a bank that wants to find out which of its customers can be considered “most valuable,” or in other words, least likely to leave while also producing the most revenue for the bank. In this case, two predictive models can be built:

 

  • A classification model that can be used to predict a customer's likelihood to leave (also known as attrition), and
  • A regression model that can be used to predict revenue per customer.

 

The bank can use the results from these two models to make conclusions and apply a strategy. Depending on the model outputs, some of the possible strategies include:

 

  • Deciding to do nothing for the customers that have a low or medium spend and a low or medium risk of attrition. Retaining these customers does not provide as much value to the bank as other customer types.
  • Deciding to apply a low-cost retention strategy for customers that have a medium spend and medium risk of attrition.
  • Deciding to put most of the bank’s efforts into retaining customers that have a high spend and high risk of attrition.

 

Regardless of the choice the bank makes in this example, predictive modeling has provided them a range of options by which to potentially increase the value of their business and customer satisfaction.

Classification models and regression models are two ways of using machine learning to answer many different types of questions. While deploying a new machine learning-based tool within your organization can be a difficult task, framing your questions with the knowledge of how machine learning works can bring you valuable answers and big benefits.

About the author

Claudia is a data scientist, consultant and trainer. She is the CEO of Edlitera, a data science and machine learning training and consulting company helping teams and businesses futureproof themselves and turn their data into profits.

Before Edlitera, Claudia taught Computer Science at Harvard, and worked in biotech (Qiagen), marketing tech (ZoomInfo), and ecommerce (Wayfair). Claudia earned her degree in Economics from Yale, with a focus on Statistics and Computer Science.