Wonderful Wines of the World (WWW) is a ten-year-old enterprise that seeks out small, unique wineries worldwide and brings its wines to its customers. Its mission is to delight its customers with well-made, unique, and exciting wines that would never travel far beyond their points of origin. WWW sells wines through catalogs (electronic and physical), websites, mobile apps, and ten small stores in major cities around the USA. Customers can purchase at the stores, by telephone (after looking at the catalog), or through the website/mobile app.
Through aggressive promotion in wine and food magazines, WWW now has 350,000 customers in its database. Most customers are highly involved in wine, entertain frequently, and have sufficient money to indulge their passion for wine. WWW sometimes offers wine accessories as well – wine racks, cork extractors, etc.
WWW is trying to make use of the database it started about five years ago. So far, it has simply mass-marketed everything. Now, WWW wants to "get smart" about its database, and start differentiating customers. WWW has provided one random sample of its customers from its active database. These customers have purchased something from WWW in the past 24 months (after 24 months with no purchase, a person is eliminated from the active database).
The dataset used for this project contains the following variables:
Custid: Customer ID, uniqueDayswus: Number of days as a customerAge: Age of the customerEducation: Highest academic degree earnedIncome: Household net incomeKidhome: 1=has child under [0y-12y] living at homeTeenhome: 1=has teen [13y-19y] living at homeFreq: Number of purchases in past 24mRecency: Number of days since last purchaseMonetary: Total sales to this person in 24mLTV: Customer lifetime value (derived variable)Perdeal: % Purchases bought on discountDryred: % Of wines that were dry red winesSweetred: % Sweet or semi-dry redsDrywh: % Dry white winesSweetwh: % Sweet or semi-dry white winesDessert: % Dessert wines (port, sherry, etc.)Exotic: % Very unusual winesWebPurchase: % Of purchases made on website/appWebVisit: Average visits to website/app per monthExpressedPreference: Explicitly preferred line of business (LOB)NPS: Net promoter scoreHumid: Target variable, binary
Note: DRYRED + SWEETRED + DRYWH + SWEETWH + DESSERT = 100%
The prediction task was performed using two machine learning algorithms: Neural networks and Decision trees. Various features were utilized to train the models and predict the "Humid" target column.
Using the best-performing model, I applied it to unseen data to predict. The results of the predicted unseen data are used to calculate Return On Investment for the marketing campaigns.
The following dependencies are required to run the code:
- SAS Enterprise Miner 15.2 version
Neural Network, Number of Hidden Units 10, is selected as the best model.