Supervised learning involves teaching an algorithm to come to a specific conclusion based on historical data. For example, if the question is “Will this customer churn?” an analyst can look at historical data about who has churned in the past and train an algorithm to determine which customers are most likely to churn looks like based on that data. In a nutshell, an analyst creates a training data set with a known outcome (i.e. churn or not-churn) which the algorithm then uses to create a predictive model based on historical data.
Unsupervised learning involves teaching an algorithm to look for similarities or patterns in data and group things together based on that information without being taught what to look for. For example, a streaming media platform might use unsupervised learning to group users together based on similarities in viewing behavior. The streaming service may then use these clusters or segments to provide a recommendation on what to watch next.
Three types of algorithms used for predictive modeling are:
Classification: a supervised algorithm that predicts a category or “class label” based on historical data. Example: An email client labeling an email as “spam” based on a classification algorithm that considers past attributes of spam emails.
Regression: a supervised algorithm that predicts a value or number based on historical data. Example: Based on the location, size and other factors, a regression algorithm can predict the value of a house.
Clustering: an unsupervised algorithm that sorts data into groups based on similar patterns and characteristics. Example: An e-commerce website can use a clustering algorithm to sort customers based on browsing and purchasing history to help inform marketing strategy.
Predict which customers are likely to churn within a given period so you can take action to prevent the loss of valuable customers
Categorize customers into predefined groups (aka segments) based on patterns to learn more about them
Predict which patients are likely to miss their appointments so you can improve clinicians’ productivity by ensuring minimal “downtime” due to no-shows
Predict which patients are likely to be unsatisfied and why; use that information to determine how to improve patient satisfaction
Predict which policyholders are likely to lapse and come up with a strategy to increase retention
Predict which claims are likely to be successfully subrogated so you can focus efforts on high-potential claims and optimize recovery of loss payments
Predict which survey recipients are likely to respond
Predict which customers are likely to respond to campaign messages and prioritize outreach to those customers
Predict which potential customers are likely to respond and prioritize contacting them
Predict which other products customers are likely to buy so you can focus cross-sell and upsell efforts