In this two-part series on data, we discuss small but useful data projects that you can get up and running in a few weeks' time. In this second case, we explore how you can get to know your customers better. Also find out how you can upgrade your website architecture through data.
Now on to the subject at hand:
Understanding the behavior of your customers
Let’s say you want to cleverly organize your email marketing by targeting specific customer segments with tailor made content. To achieve this, you have to know which groups your customers can be categorized in. You can identify these groups by analyzing actual behavior on your website or web shop.
When you understand the different customer segments, you can target customers that are interested in specifics, like a new feature, a product, or a promotion you'll run. This will help you waste less resources on customers that are not interested in this specific content. More so, you will and increase the conversion rate of your content. Ultimately, your communication will be more efficient.
Segmentation results can also provide you with insights that help you improve certain key performance indicators (KPI's). For instance: you could find out who you should target if you want to address one-time buyers.
Customer segmentation is all about discovery. And - we're sure you’ll agree - discovery is fun!
It’s really exciting to get your hands on a new customer data set and start the pattern hunt. So let’s dive in
Non-negative matrix factorization
Customer segmentation relies on clustering techniques that look for patterns in customer behavior. Stuff like which products the customers buy and the pages they visit.
In this example we'll look at purchase history. Let us introduce you to an interesting clustering method: non-negative matrix factorization (NMF). This method is particularly interesting in cases where each customer buys only a few products from a large product offering.
Let's get a bit more technical. NMF is a type of dimensionality reduction in which the input data is decomposed into a weights matrix and components matrix. The result is an approximation of the original matrix when multiplied. Thanks to the non-negative values, you can interpret the segmentation as how much a certain customer belongs to each segment (weights matrix) and how much each segment is dominated by certain products (components matrix). The latter can be interpreted as feature importance.
"Customer segmentation relies on clustering techniques that look for patterns in customer behavior"
What could these results mean to your business?
Analyzing the results of NMF might show you cross-selling opportunities. For example, in a supermarket the segment averages could indicate that one of the segments buys both fresh as well as frozen products, and another group buys milk combined with grocery products. Or you might identify customers who buy very little of a certain product in contrast to another similar group. Perhaps they would also be interested but don’t know about that product’s existence.
The NMF segmentation can easily be extended to actual recommendations. Multiplying the components and weights matrices results in a reconstructed sales matrix in which the 0’s (or non-existing purchases) are replaced by values based on customers with similar purchase histories. Finally, a special type of segmentation that is often overlooked, is cohort analysis, most often based on time cohorts. With this method you can investigate the effect of certain experiments or campaigns in terms of – for example - customer retention or amount spent. If a campaign was successful, it will show up as higher numbers for the corresponding period in the cohort heatmap, which is a standard way to visualize a cohort analysis.
Does experimenting like this sound fun to you?
Then you should take a look at our blog post about website experiments!