Let's dive into the recent project with Plan!

Lately, Plan Sweden mentioned that they wished to adapt better their ways of interacting with donors to different types of donors: while some donors enjoy frequent updates via digital channels, others prefer a lower interaction intensity, primarily through analog means of communication. Yet, the question is, which categories exist and which donors belong to which category? 

The solution we discussed with our partners from Plan was based on a hybrid approach.

Building on quantitative interaction data, we fed a machine learning model that came up with several donor clusters (these clusters were our hypothesis for the “types of donors” that we referred to above. In addition, we gathered qualitative information by interviewing donors that were categorized into these defined clusters. By doing so, we better understood these donors’ preferences and put a face on the typical donor in each segment (so-called “personas”). 

As a first step, we developed the machine learning model. It is critical to feed it with the right, relevant features as input for the segmentation model. To select these features, you need to ask what is useful and meaningful to the use cases you have in mind. This is crucial because all chosen features will affect the result, irrespective of actual influence. For example, suppose “donor age” is selected as a feature in the model. In that case, the segmentation will, to some degree, reflect donor age, even if donor age is entirely irrelevant for the use cases in mind.  

To define these input features, we relied on a proven internal framework, BearingPoint’s Value-Driven Design framework. Together with our colleagues from Plan Sweden, we  

1. Identified behavioral features of interest,  
2. Looked at what actions the segmentation will enable,  
3. Determined what insight is necessary to inform the action making, and  
4. Looked at what data is available to draw conclusions and reach insights.  

We created a long list of potential features that we then extracted from Plan’s databases to perform feature analysis. With this kind of analysis, we determined the features to end up in the final model. The machine learning model we fed was a so-called k-means model. This is an easy-to-understand statistical method to build clusters (segments). Its output helps categorize users into segments – fairly straightforward. Based on this method, we identified four segments with different user types. 

Then the qualitative step followed. We interviewed close to 100 donors. The goal of these interviews was to validate segment donors’ preferences and describe these in more detail. As interview partners, we selected donors to represent the defined clusters evenly. With the feedback from these structured interviews, we derived qualitative descriptions for each segment and, based on these, developed personas¹.

¹ Personas are fictional characters that represent different donors types and shared characteristics, preferences and/or behaviors. Personas put a “face” on each segment, improve the understanding of typical characteristics, and, thus, simplify communication. 

These days, Plan Sweden relies on the build segments to differentiate communications and interactions with donors regarding the preferred type and intensity of communication. In the long-term, better understanding and addressing donor preferences holds the potential to reduce churn and improve overall donations. Our tool can also identify donors who are more prone to take action on upselling campaigns, raising ROI on the sometimes costly sales campaigns. 

"We, as a children’s rights organization, are completely dependent on our donors, and it is of the utmost importance to understand them. The donor segmentation helps us communicate and interact with our donors in the right way, which secures current and future support for vulnerable children."

Malin Wessman, Plan International