Improving Starbucks Offer Targeting with Offer Success Prediction

Context

This simulated Starbucks Rewards App dataset was a capstone project option by Udacity Data Scientist Nanodegree with an open-ended problem statement to derive business value for Starbucks.

Here, it is used to build a model that can better rank which offers to serve to each customer. We aim to address the twin problems of:

1. Wastage - Customers are sent (and therefore use) offers even though they would made a purchase without one
2. Irrelevance - Irrelevant offers are sent which customers tune out, leading to missed conversion opportunities

Objective

• Understand current offer redemption trends and highlight improvement areas
  • See 02_Exploratory_Data_Analysis.ipynb
• Build a framework to rank offers to serve to customers by:
  • Tuning and select a classifier that predicts redemption
  • Tuning and selecting a regressor that predicts expected spend if redeemed
  • Combine both to output an understandable ranking of offers
  • See 03_Redemption_and_Expected_Spend_Prediction.ipynb

Dataset

This dataset is available from Kaggle (https://www.kaggle.com/ihormuliar/starbucks-customer-data). Data is simulated but closely mimics customer behavior over the Starbucks Rewards App.

It consists of 3 csv files:

1. portfolio.csv: 10 entries with offer configuration information
2. profile.csv: 17,000 entries with customer demographic information
3. transcript.csv: 306,534 rows of 4 types of interactions; offer received, offer viewed, offer success and transactions

Exploratory Data Analysis Highlights

• Success Rate: Only about 42% of all promotions sent out are ultimately redeemed on purpose ("success").
• Wastage: 13% of customers do not even see the promotion and yet redeemed it suggesting that the promotion was unnecessary and value was lost
• Frequency: Customers received on average 4 offers a month of varying configurations
• Offer Success differs most significantly by
  • Gender: There is higher success with females compared to males
  • Duration of being a member: Long term members (>500 days) have higher success
  • Social Media: There is higher success if the offer is on social media as well
• Spending differs most significantly by:
  • Age: Older individuals spend more
  • Income: High-income individuals spend more

Packages used: pandas, numpy, matplotlib, seaborn

Model Building Approach

Predict Offer Success

• 4 classifiers were tuned and trained (Logistic Regression, Support Vector Machine, Random Forest, XGBoost)
• XGBoost had the best performance, with 73% precision (this is a significant improvement from the 42% business success rate above)

Predict Expected Spend

• 3 regressors were tuned and trained (Linear Regression, Random Forest, K-Nearest Neighbours)
• Random Forest had the best performance, with 0.5 r-squared.

Offer Ranking

1. Demographic and promotions data first fed into the classifier to get P(Offer Success)
2. The same features are then fed to the regressor to get Expected Spend
3. The promotions are then ranked by P(Offer Success) x Expected Spend

Packages used: sklearn, xgboost

Files

• 01_Data_Preprocessing.ipynb: Data formatting, filtering and grouping
• 02_Exploratory_Data_Analysis.ipynb:
• 03_Redemption_and_Expected_Spend_Prediction.ipynb: Classifier & regressor tuning, selection & combined expected value framework
• data: 3 original CSV files, 1 preprocessed CSV file and 1 example to demonstrate model application
• images: Supporting images

Collaborators

This project was co-written with Toh Pin Xuan (@tohpinxuan).

Photo by Fahmi Fakhrudin from Unsplash