gscottstukey / AdVantage

AdVantage

A machine learning mobile ad monetization optimizer.

Overview:

• A data product that predicts the mobile ad revenue prior to receiving the ad from ad networks.
  
• Using Ad Vantage, mobile apps or ad exchanges can optimize revenue.
  
• Advertisers can better understand their targeting demographics, publisher content relevance and user behavior.

Results, Accuracy and Cross-Validation:

• Prediction accuracy: 79%.
  
• Random Forest ROC Plot AUC: 0.85.
  
• Logistic Regression ROC Plot AUC: 0.76.
  
• Naive Bayes Gaussian AUC: 0.60.
  
• Naive Bayes Multinomial AUC: 0.73.
  
• Cross-Validation F1 Score Random Forest: 0.89.
  
• Revenue Lift from "Random" baseline: 109%.
  

Data:

Logs of raw ad tags received from ad networks to classify five targeted revenue buckets.

Analysis and Modeling:

• The Classifiers were Random Forest, Logistic Regression (One-vs-All), Naive Bayes (One-vs-All).
  
• Tuned Classifiers with Grid Search, conducted Feature Engineering and cross-validation.
  

On the Roadmap:

• A/B testing to verify revenue lift; or lift chart.
• Feature extraction, create device user session feature, behavior pattern.
• Core revenue group: top 20% apps, devices and locations.
• Train on more time-consuming models to improve accuracy.
• Build scalability.

Project Details:

• Code files at:
  project_code
• Project Presentation:
  Jun Zhang presso mobile ads.pdf
• Some Visualization and more project details to come at:
  http://junkateannzhang.herokuapp.com/