Segment 2,200+ customers to reframe targeted re-engagement strategies

Role: Data Analyst / UX Researcher

Approach

Goal:

Identify the best model to create customer segments using the given customer dataset. Once we identify the best customer segments, we can prioritize investment in communication and advertising to achieve the highest customer engagement

Exploratory Data Analysis Results:

Data Breakdown

Data Description

Demographics Trends

Buying Behavior Trends

We see a large number of purchases in lower quantities for sweets, fish, and fruits.

But on average, larger quantities are purchased for meats, wines, and gold products.

Purchasing Behavior Trends

Sales Engagement Trends

Model Selection

Since the goal of the research was to identify customer groups to better target them in the context of a product campaign, and the data provided were customer demographics and engagement, the best machine learning clustering algorithms for this are those that group unlabeled data into clear clusters. This includes models such as K-Means and DBSCAN. I decided to test multiple models to see the performance of each and select the best-performing one, the one that creates the clearest clusters and profiles,  to move forward with.

The Solution Formulation

1. Apply Principal Component Analysis (PCA) to reduce the dataset's dimensionality while retaining most of its variation, making underlying clusters easier to identify.

2. Then, we compare K-means and K-medoids using visualizations and silhouette scores.

3. Next, we apply Hierarchical clustering, DBSCAN, and GMM to gain perspectives.

4. Review models to ensure consistent themes or improve performance.

K-Means and K-Medoids both cluster data based on similar features, while K-Means approximates a center for each group. K-Medoids selects a specific point in the dataset to cluster around, avoiding skewing from outlier values that could drastically change the shape or position of the clusters.

Insights:

We can see that K-Means does a better job here of creating separate clusters, while K-Medoids has some overlap.

DBSCAN groups data into clusters based on the density of neighbors, allowing for more variation in cluster sizes and avoiding skewing by outliers that K-Means and K-Medoids can struggle with.  

Hierarchical clustering allows me to create a decision tree by grouping the data by different features in the dataset until they can be split into increasingly distinct subgroups. The higher the vertical line, the more stable and reliable the groupings are for understanding the data.

Insights:

The visualizations show that each clustering method separates the data into two primary groups, although DBSCAN and K-Medoids exhibit greater overlap between clusters. Based on the visual separation alone, K-Means and Hierarchical Clustering appear to produce the clearest groupings, making them strong candidates for developing segment profiles.

Rather than relying solely on visual inspection, I also compared the models using silhouette scores. This metric measures how similar data points are to others within the same cluster relative to those in neighboring clusters, providing a quantitative assessment of cluster quality and separation.

Silhouette Scores

This means the dataset shows shared characteristics between the groups, which I need to be mindful of when creating customer segment profiles and recommendations.

From all these models’ scores, I was confident moving forward with K-Means, as it had a slightly higher Silhouette score while keeping overlap among the groups in mind.

Customer Segmentation Profiles

Budget-Conscious Browsers

Demographic

On average, their household income is around $40,000.00.

They are also more likely to have children at home and do more shopping online.

Buying Behaviors

They are more likely to have engaged recently with the company.

2-3X as likely to engage via the company’s website as compared to other channels.

Product Purchases

They make about half as many purchases across products as the High-Value Engagers.

Sales Engagements

Less likely to actually make purchases, and when they do, they spend less and are more likely not to use discounts. (4X less spend per purchase compared to High-Value Engagers.)

Recommendations

Web Experience optimization

With these customers shopping and spending more time online, we should make sure our products in that space align well within these customers’ price pain point.

Product Specific Marketing

To drive greater engagement, we can tailor discounts and product offerings to better meet the needs of families and multi-person households.

High-Value Engagers

Demographics

They make around $75,000.00 in their household income on average.

They are less likely to have children than our Budget-Conscious Browsers.

Buying Behavior

They make more purchases per day and purchase 2X as many goods as our other group.

While they are more likely to shop in store they are more likely to make purchases across all channels of engagement.

Their spend per purchase is 4 times as large as our Budget-Conscious Browser.

Product Purchases

They are more likely to purchase goods from all categories but higher for meat and wines

Sales Engagement

They are more likely to engage with Campaign efforts at 4 times the rate as our Budget-Conscious Browser.

Recommendations

Campaign Engagements

Make sure we get a large volume of campaigns out to these customers, as they are more likely to engage

In Store Sales

As these customers make more purchases in stores and use discounts, we can make sure to make these discounts ( coupons, sales, etc) visually prominent in the physical store experience

'

Cross Group Recommendations

Further Research Suggestion and Reflection