Data driven customer segmentation: Our approach to actionable segmentation

Our approach to actionable segmentation

Personalizing customer relationships is no longer a choice, but a prerequisite that significantly impacts the value of customers throughout their purchase cycle: from acquisition to recommendation and retention. To implement this personalization strategy, marketers have various tools and techniques at their disposal, based on a detailed knowledge of customer behavior, acquired through the use of customer data. Segmentation is an essential, simple and powerful tool. It is the first step in building a personalized customer strategy, allowing differentiated communication with each group of customers with similar behaviors.

In this article, we are interested in the contribution of data science to this classical technique and how our methodological approach allows us to objectify the segmentation while integrating business constraints and knowledge. We present our approach to reconcile business relevance and statistical robustness.

Data-driven marketing segmentation: an indispensable tool

Segmentation is an operation that consists of grouping customers with similar characteristics in order to carry out appropriate and differentiated actions on each group. Marketing segmentations have long been based on beliefs about customer profiles and behavior or on quantitative and qualitative studies. But data science can objectify these analyses, by exploiting customer behavior and profile data. These data-driven segmentations will take into account different types of data: demographic data (customer profile), geographical data, behavioral data (purchases, browsing, etc.) and customer preferences data. The criteria that can be used in a segmentation are multiple and depend on the company’s sector of activity but also on the final objective of the segmentation.

The implementation of data driven marketing segmentation is based on clustering methods. The objective is to group together customers sharing similar attributes within the same cluster. In addition to this purely statistical work, the segmentation must integrate sector specificities and business constraints in order to be relevant and actionable. The data scientist’s work does not stop once the clustering is done but carries on to deliver interpretable insights. Indeed, the implementation of relevant activation strategies requires a strong understanding of customers clustering criteria along with their profiles’ differences. In order to allow marketers to use segmentation, data scientists must interpret the algorithm and make it easily understandable, even if it means losing some of its discriminatory quality.

The segmentation process will be accompanied by the implementation of differentiated communication actions on each segment, allowing for preferences’ adaptation to all groups characteristics. This personalization will have a significant impact on customer behavior (in terms of purchase, re-purchase and recommendation) but also in terms of customer value. It will enable the delivery of an appropriate message through a fine-tuned understanding of the customers’ group needs, which will consequently lead to customer experience improvement, and therefore, better customer retention. In addition, monitoring the segmentation over time will help identify emerging trends.

Set up a data driven segmentation

As is often the case in Data Science, there is no magic formula but rather a set of techniques that are appropriate for specific use cases. Nevertheless, we will present a generic approach that will allow us to perform a robust segmentation while integrating business constraints and rules.

1. Definition of the segmentation objective and listing business constraints

The first step is to define beforehand the objective and business specificities in order to determine the set of constraints to be integrated in our segmentation work. Here is an example of the list of questions needed to create an actionable segmentation:

• What is the objective of the segmentation: To develop loyalty, customer value…?
• Do you want to develop a new segmentation or refine an existing one?
• Are there priority groups of customers to isolate beforhand (e.g. new customers)?
• What is the desired level of granularity?
• Are there any segmentation criteria that are not (or no longer) relevant in the short term (e.g. change in product line or price)?

2. Quality and exploration

We won’t go into detail about the data collection step, even though it is essential to have a complete vision of customer behavior and appetites. Once this part is over, we must look for potential quality problems, data anomalies and correct them with the appropriate techniques. This step is the cornerstone of the marketing strategy that will be put in place. Indeed, if the customer data is incomplete or of poor quality (e.g. presence of duplicates, nan values), the results will not meet our initial expectations nor the clients needs. As a consequence, interpretability will be biased.

After the data sanity check, we move forward to data exploration. Exploratory data analysis (or EDA) helps us calculate the first indicators describing customer behavior throughout first univariate and then multivariate analyses. This is a key step to better understand customer behavior and also enrich our statistical segmentation with preconceived business ideas about customer behavior which will set the first brick for the activation strategy.

3. Segmentation

The sole purpose of the segmentation step is to create homogeneous groups of people. We can divide this step into three phases: First, we start with data preprocessing in order to optimize our use of explanatory variables as it is necessary to reprocess continuous and categorical variables accordingly.

Secondly, upon data processing completion, two choices are possible:

• Use clustering methods without dimension reduction in order to keep all the explanatory variables in the segmentation
• Add a dimension reduction step (PCA) to segment customers with fewer explanatory variables.

In this last case, which remains optional, the goal is to reduce the number of explanatory variables as well as the collinearity between them. Beside the fact that these techniques provide us with runtime optimization, we should keep in mind that they also create new variables that might be uninterpretable in business terms. Finally, the best algorithm to perform the segmentation must be chosen. We generally compare two approaches: ACH and KMeans, both in terms of robustness and model quality but also in terms of business representation of the results.

4. Interpretation of segmentation results

Once the statistical segmentation has been carried out, it must be interpreted in order to provide marketers with the elements they need to implement a differentiated marketing strategy. The objective is to qualify each segment and understand the common characteristics shared by customers belonging to the same cluster. In this context, several methods exist and can even be combined: For example, the analysis of discriminating variables, the construction of decision trees to reconstruct our clusters.

Analyzing the temporal stability of the segmentation alongside the intra-cluster evolutions over time will provide us with the key elements to define the customer’s natural evolutions and identify the virtuous paths to be pushed during the activation program.

At the end of this step, we will have the criteria upon which the clustering is based, allowing us both to define the typical customer profiles for each segment and also to build the ones that will be used for the activation strategy. Furthermore, cross-analyzing segments’ properties with the initial exploratory analysis will help us retrieve marketing issues within each group as well as define the key moments of the purchase cycle. Thus, we will reach the ultimate step of our actionable segmentation by setting up an adapted and differentiated strategy based on quantitative elements combined with business knowledge.

In this article, we have detailed a data-driven customer strategy approach based on an interpretable statistical segmentation that takes into account business objectives and constraints. In a next article we will propose a review of the clustering and interpretation algorithms that we can use in these techniques.