Starbucks offers: Advanced customer segmentation with Python
How to use cluster analysis to target marketing outreach
This is a Udacity Data Science Nanodegree Capstone project.
A small startup can afford to target users based on broad-stroke rules and rough demographics.
Once a company grows to the size of Starbucks, with millions of daily customers, and $1.6B in credit stored on loyalty cards, they have got to graduate to a more sophisticated method to target their marketing.
One such approach, cluster analysis, uses mathematical models to discover groups of similar customers based on variations in their demographics, purchasing habits, and other characteristics.
Below, I will explore a customer transaction and marketing offer dataset graciously provided by Starbucks.
I will then use Principal Component Analysis (PCA) and the k-means unsupervised Machine Learning algorithm to group these customers into clusters that can be used to automate an effective outreach campaign.
All code behind this blog post is available on my GitHub.
One month of transaction data
The dataset includes one month of simulated customer data, including their purchasing habits, and interactions with promotional offers.
Each person in the simulation has hidden traits that influence their purchasing patterns, and that are associated with their observable traits.
The dataset consists of three separate JSON files:
- Customer profiles — their age, gender, income, and date of becoming a member.
- Portfolio — Offers sent during the 30-day test period, via web, email, mobile or social media channels, or a combination thereof. The offers have varying levels of difficulty (minimum spend) and reward, and fall into one of three categories: Discount, Buy-one-get-one (BOGO), Informational
- Transcript — A list of offer interactions (receive/view/complete), and all other transactions during the test period.
Starbucks serves tens of millions of customers that cut across demographics, and have very different tastes and purchasing habits.
To make the most of our outreach, we need a way to take this complexity, and automatically segment customers into groups that respond best to a particular marketing campaign.
Unlike supervised Machine Learning algorithms, unsupervised clustering does not have clearly defined metrics for the optimal parameters, or number of clusters. Below, I will use the elbow method and silhouette coefficient to validate the clustering algorithm’s performance, and choose the best number of segments for our data.
I then analyze the view rate (% of customers who viewed a campaign after receiving it) and conversion rate (% of customers who complete an offer after viewing it) for the different clusters as a way to assess how useful these segments will be for our business.
Before I could visualize and model the data, I’ve had to do some preprocessing both outside, and in Python.
Among others, I have:
- Removed empty lines in transcript.json using search \n\n & replace \n in Visual Studio Code
- Imputed empty income values with the mean ($65,404), and added a separate feature that tracks missing income values with 1s and 0s.
- Engineered a new feature for the year when the user became a member
- One-hot-encoded channels using the MultiLabelBinarizer
- One-hot-encoded offer types, genders, years joined and event types using get_dummies
- Dropped age outliers (a number of outlier customers had their age set to 118, and were missing data for several of the other fields)
- Engineered first receipt, first view and first completion time features (a customer can receive and interact with the same offer multiple times)
- Dropped misattributions (completion without view, completion before view, or view before receipt)
- Calculated RFM Recency and Frequency scores— a common method used for analyzing customer value in retail and e-commerce
- Engineered view and conversion rate features for each offer type
- Merged all data into one dataframe grouped by customers, including means and sums for all available data, as well as additional columns for the average number of exposures per offer-type.
The dataset 306,534 events related to 17,000 customers (14,808 after data cleanup) and 10 event types over the course of a 30-day experiment.
The majority of customers in the dataset are male.
The mean age across all customer groups, after removing outliers over 99, is 53 years. Male customers in the dataset tend to be younger than this average.
Incomes range from $30,000 to $120,000, with a mean of $61,800. Female customers tend to have higher incomes than male customers, likely correlated with their higher average age.
The average transaction value is just $14, but there are long tail transactions of up to $1062.
Looking at conversion funnels, BOGO offers have a higher view rate, but a lower conversion rate than discount offers.
1. Feature scaling
Both the PCA and k-means algorithms which we will use below are sensitive to the relative scale of the data.
For example, our boolean columns range from 0 to 1, whereas the income column ranges from 30,000 to 120,000, a different magnitude which would negatively affect the clustering.
To solve this problem, I’ve used StandardScaler to transform data such that its distribution will have a mean value of 0, and a standard deviation of 1.
2. Dimensionality reduction
The k-means algorithm is both more effective and more efficient with a small number of dimensions, that is, the number of features used to predict the right cluster for each customer.
To reduce dimensionality, I’ve used Principal Component Analysis (PCA) — a method which identifies variables that are responsible for most of the variance in the data.
I have first used the common k-means algorithm to classify the data, settling on four clusters based on the elbow method and silhouette score heuristics.
k-means, with n_clusters=4:
I have then experimented with DBSCAN and OPTICS, two density-based clustering algorithms.
For DBSCAN, I identified 2.5 as the optimal value for the eps parameter, using the elbow method.
For both DBSCAN and OPTICS, I then experimented with a variety of min_samples values that would generate a reasonable number of well-differentiated clusters.
DBSCAN, with eps=2.5 and min_samples=150:
OPTICS with min_samples=40:
Evaluation and validation
Although the density-based algorithms may appear to perform better in the charts above, the clusters generated using k-means show more distinct characteristics that make sense in our business context.
The main problem appears to be that DBSCAN and OPTICS overemphasize the gender and membership year of the customers, as those variables are more densely clustered.
The resulting data shows minimal variance in our view rate and conversion rate metrics, and is therefore not actionable in our marketing.
Compare this to the clusters generated using k-means, well differentiated not just in their demographics, but also conversion rates for each individual offer type:
Above, we can immediately identify four distinct segments with clear business implications:
Customers in this segment receive regular BOGO offers, and practically no discount offers. These BOGO offers involve more valuable rewards than for customers in other segments.
Their frequency and average order value are not unusual, which suggests these customers are conditioned to BOGOs, and we might have to continue sending them regular offers to keep their patronage.
BOGOs convert really well with customers in this segment, so this is a great lever in times when we need to quickly generate additional sales.
Customers in this segment receive a higher than average number of offers, and convert really well for both BOGOs and discounts. Demographically, a higher than average share of these customers selected their gender as Other.
Their average order value is not unusual, in line with the average, but their frequency is above average, probably as a result of the regular offers they receive and act on.
This is another segment we can target to quickly generate additional sales.
Customers in this segment receive no BOGO offers. They do get occasional discount offers, on which they convert about average, as well as slightly more informational messages than other customers.
These customers have about average frequency and average order value, and would likely continue to frequent Starbucks even if we stopped sending them offers.
Customers in this segment receive regular offers, which they open, but never convert.
Demographically, they are predominantly male, and lower than average income. They also visit Starbucks less frequently, and make smaller average purchases.
Given the low LTV and low conversion rates for this group, we may be best to avoid targeting them in our marketing.
This was a complex but fascinating project.
As usual in data science, cleaning and feature engineering took 80% of the time, but the resulting clustering was well worth the effort.
We’ve identified four segments showcasing distinct purchasing habits and reactions to marketing offers.
Most importantly, we’ve identified an entire segment of subpar targets that we can exclude in our paid marketing campaigns to optimize our Customer Acquisition Cost.
Many challenges in working with this dataset resulted from repeat exposures to the same offer over different channels, and imperfect conversion attribution. In future experiments, it would be desirable to generate more accurate data on the source of each conversion, and confirm completion through coupon codes or a separate redemption mechanism in the Starbucks Rewards app.
Additionally, many customers had missing profile fields. Understandably, we cannot force members to respond to some of these questions for ethical and legal reasons.
Above, I have imputed missing income values with the mean, and engineered a separate feature tracking customers who did not respond to this question. This approach could be further improved by imputing the data using a supervised Machine Learning algorithm (predicting income based on the other demographic traits), or using the mean income of residents in the customer’s neighbourhood (not provided in this dataset).
It would also be fascinating to explore more of the data, and segment customers further by product line. For example, many Starbucks BOGO offers involve new and seasonal drinks, which may receive very different reactions depending on how conservative or adventurous is the recipient of the offer.