Upgrade to Pro — share decks privately, control downloads, hide ads and more …

ML for Marketing

ML for Marketing

We built ML models to optimize revenue through marketing by predicting whether there will be coupon redemption in a transaction or not.

Bdb3e87c8ff8876185636fc58269e896?s=128

Shanelle Recheta

March 18, 2021
Tweet

More Decks by Shanelle Recheta

Other Decks in Programming

Transcript

  1. Coupon Redemption team.fit()

  2. Grocery at your fingertips

  3. 1,116 100 out of coupons redeemed

  4. What will work? Who to target in their specific campaigns?

    What is working? Which campaigns are likely to succeed? What has worked? Which coupons should be revisited?
  5. Data-driven Marketing Increase customer loyalty Increase customer engagement Increase profit

  6. Customer-first Analytics • Profiling the customers through K-means Clustering •

    Observe customer behavior through Aggregation of features • Comparing the performance of campaigns through descriptive stats
  7. Opportunity to Market: Purchasing Power Get median purchase duration between

    customers Recommend a coupon at about the same time they will be purchasing again Make the product stickier
  8. Underlying Magic: Feature Extraction Customer-centric variables Research-based (domain knowledge) Time-centric

    and Categorical variables ML-first EDA (getting high variance trees from Random Forest)
  9. Underlying Magic: Running K-means TWICE

  10. Underlying Magic Strong Positive Correlation • brand vs item type

    ◦ e.g. Packaged Meat • brand vs campaign release date ◦ Day of week ◦ Month (seasonality)
  11. Challenges Encountered • Integrating all datasets • Feature Extraction ◦

    Effectively encoding categorical columns ◦ Clustering and extracting features from what the trees say have highest feature importance • Dealing with missing variables ◦ KNN impute • High memory usage: Large number of rows = long processing time ◦ Used SQL for queries
  12. Challenges Encountered • Models tested: Random Forest, XGBoost, fast.ai ◦

    bootstrapped significantly: bagging principle ◦ Fast iteration: we didn’t have a model that went through the entire dataset ◦ All models were bootstrapped instances that were averaged out ◦ fast.ai to ensure trees weren’t correlated to maximize the feature insights acquired • Overfitting ◦ L1 regularization does the trick
  13. Insights and Recommendations Factors affecting Success of Campaign • Day

    of the week that promos come out have relatively significant effect purchases and campaign came in • There are shoppers that shop heavily on specific days • Customers buy the same way • Brand type released on certain dates (start of week, seasonality) • Coupons that can be used for many items are redeemed more often
  14. Future Work • Integrate Brand-level Categorical Variables • Product Category

    Clustering: Looking at Price Sensitivities across Brands and Categories • Item code - Categories - Brand: Mapping and Analysis • Matrix Factorization for Coupon Recommendation • Collect data on profit per item (sales =/= profit) to effectively measure profitability of the marketing campaign
  15. team.fit() Alec Wang Tech Maven Daryl Collantes Designer Shane Recheta

    Project Leader
  16. Thank you! team.fit()