The answer is not usually as simple as “let’s use the algorithm I read about in this magazine article!” Although that is occasionally the answer.
To determine the right machine learning algorithm for your business needs, you should typically consider:
- The data structure of what is being predicted
- The data structure of the predictors
- The kind of insights needed to drive changes in your business
What various algorithms can do
Suppose a business manager wants to predict which customers are going to cancel their account in the next 28 days so they can prepare counter offers in advance. They need to generate a yes/no prediction on call-to-cancel. And the modeling dataset must include descriptive information about account holders and their usage so the business can tailor an offer to match their account profile.
If you have data about the number of transactions each customer made during each week of the past 12 weeks, you could use a Logistic Regression algorithm. This algorithm predicts a yes/no answer from a set of numeric inputs, and gives an equation from which you can infer the likelihood of certain customer behavior.
Another choice might be a Decision Tree that produces a yes/no prediction and a set of descriptive rules or profiles for each prediction.
For machine learning, the key for any of these algorithms is how you stage the data. The business information needs to be presented to the computer in ways that enable the system to learn which patterns of past customer behavior are associated with each result. Sometimes this is referred to as feature selection. But it’s more than picking a few features from a list of columns in a database.
For example, suppose you have a theory that customers who want to cancel are the ones whose usage is declining. If so, then you need to set up the model to look at a period of time over which to examine customer activity, determine their baseline usage, and then check for changes (increases and decreases) in usage across that time span.
That data, along with other descriptive information, will help an algorithm look for subtle variations and patterns in behavior that lead to predictive behavior. Then when current customers have behavior profiles that are similar to previous customers, they can be flagged for corrective action to try to preserve the relationship and prevent attrition.
Selecting the right predictive model takes strategic data preparation and creative experimentation — and it’s worth the effort.