Understanding the characteristics of active app users

Problem statement

After a year of being in market, the user base of our loyalty app had grown dramatically and was hailed a success by the business. But we were now entering the second year and knew that to maintain high user acquisition levels, we needed to incorporate a scientific approach in our marketing efforts.

Background

With a year of data now collected on the app and its users, we wanted to learn more about our most active users. Knowing who these users were and how they were using the app would help guide marketing and product management efforts, by answering questions such as:

• What is the most popular feature used by active users, so we can promote this to prospective users?

• What type of member do active app users represent, so we can find similar members who may not have heard about the app yet?

The business defined an active user as a member who had opened the app (and logged in or were in a logged in state) at least once in the prior week / last 7 days.

Approach

(1) Hypothesise a list of attributes

As this was a relatively broad question, in order to gain meaningful and practical insight, we first needed to brainstorm a list of attributes or characteristics that could be useful, source the relevant data for these, and then overlay that to active app users.

My approach was to identify usage behaviours or characteristics that could inspire our approaches to marketing. For example, attributes that could be used to build lookalike models (which find prospects with similar characteristics, suggesting the potential to be converted into active app users).

The list of attributes I worked with included some of the following:

• Frequency of app use by week, in the last 4 weeks 

• Whether the customer had become a member recently or not

• Key app features used in the last week

• Engagement level of the member with the program (a set of already defined business segments)

• The number of loyalty partners the member had linked their card to

(2) Collect the data needed for the attributes identified

Some of these attributes existed as transformed data in our database, others were new and needed to be created for our purpose. (If these new attributes proved useful, we would also add them to our existing feature store.)

I created a table at the member / app user level, with a column for the member ID and fields for each attribute. For simplicity I took a snapshot in time to perform this first pass exploration, by using a recent complete week – making sure to choose one where no major marketing campaigns had been run (to avoid any skews in user behaviour).

I either sourced already available attributes, or transformed data into newly existing attributes myself for each member – for this week in time.

(3) Visualise results to synthesise and draw insight from the data

In presenting results back to the team, I grouped attributes into general categories:

• User demographics

• App usage behaviour

• Loyalty program engagement behaviour

Then populated each slide with a chart per attribute, breaking down the values across active app users. Displaying these in a pack, I highlighted which attributes showed larger skews or spread in the data, compared to others where active app users appeared relatively equal across all values of that attribute.

 Outcome

By dedicating time upfront to planning and brainstorming a list of new attributes to explore, this exercise led to new insights and was a key input into future data science models.

At a high level, most attributes weaved a similar story to that the business was already aware of – that members who were engaged with the loyalty program as a whole, were also active app users. However, by spending the time upfront to brainstorm and create new attributes, this exercise provided deeper insight.

The key insight was in finding a set of additional attributes which served as good candidates to include in a predictive model. Our Data Scientist later used these attributes to build an acquisition model, that scored non-app users based on their predicted propensity to download the app if we marketed to them.

When implemented a few months later, this model went on to save us tens of thousands of dollars by improving our marketing efficiency. We knew which members didn’t need an incentive to download the app, versus others who did need some incentive (in the form of a points offer) to download the app.