How I Built a Revolutionary Marketing Campaign Optimization System That Outperforms Traditional Methods by 25% - Bachelor's Thesis
Revolutionary MINLP-XGBoost marketing system achieves 25% performance boost. Portfolio showcasing advanced analytics & optimization expertise.
The Problem That Sparked My Research
Picture this: You’re a marketing manager with a $100,000 monthly budget to allocate across Facebook Ads, Google Ads, and DV360. You need to decide not just how much to spend on each platform, but also predict what results you’ll get. Traditional tools either optimize your budget OR predict performance—but never both effectively. Add in limited computer processing power, no knowledge and data to work with advanced AI systems such as Reinforcement Learning or Neural Networks, and you’ll have a headache.
This disconnect between optimization and prediction with limited resources creates a massive gap in marketing decision-making. That’s exactly the problem I set out to solve with my thesis project.
What I Built: A Game-Changing Hybrid System
I developed the first-ever proof-of-concept system that seamlessly integrates:
- MINLP (Mixed Integer Nonlinear Programming) for intelligent budget optimization
- XGBoost machine learning for accurate performance prediction
- An intuitive interface that marketing teams can actually use
all in an easy-to-develop, accessible framework, allowing marketers to take campaign decisions on the spot using information and recommendations.
The Results Speak for Themselves
- 25% performance improvement over traditional linear programming methods
- 2-3 second response time for real-time campaign planning
- 81% accuracy in conversion predictions (R² = 0.81)
- 68,178+ campaign days analyzed across 3 major advertising platforms
The Technical Innovation Behind the Magic
Why MINLP Was a Game-Changer
Most marketing optimization tools use simple linear programming, which assumes that spending twice as much will get you exactly twice the results. Anyone who’s run ads knows this isn’t reality.
My MINLP model captures the real world:
- Diminishing returns effects (that first $1000 works better than the 10th $1000)
- Nonlinear relationships between budget and outcomes
- Platform-specific constraints and behavior patterns
The mathematical proof is in the results: 13-25% better performance across different platforms compared to traditional linear models.
XGBoost: The Prediction Powerhouse
While MINLP tells you how to allocate budget optimally, XGBoost predicts what will actually happen. I trained two specialized models:
Conversion Model Performance:
- R² = 0.81 (excellent accuracy)
- MAE = 150.55
- RMSE = 389.27
- MAPE = 10.35%
Awareness Model Performance:
- R² = 0.60 (good accuracy for awareness metrics)
- MAE = 8.76
- RMSE = 21.11
- MAPE = 56.15%
Both models significantly outperformed baseline approaches (p < 0.0001), validated through rigorous 5-fold cross-validation.
The Development Journey: Challenges and Breakthroughs
Data Wrestling: 68,000+ Campaigns Don’t Organize Themselves
Working with real advertising data from Facebook Ads, Google Ads, and DV360 taught me that data preprocessing is often 60% of the battle. I had to:
- Clean and standardize metrics across different platforms
- Handle missing values and outliers intelligently
- Engineer features that capture platform-specific behaviors
- Balance the dataset to avoid bias toward any single platform
The Integration Challenge
Building two high-performing models is one thing. Making them work together seamlessly is another. I designed a sequential hybrid system where:
- MINLP optimizes budget allocation based on user inputs
- XGBoost predicts the expected outcomes of that optimization
- A sanity check system validates results and prevents edge cases
- The GUI presents everything in marketer-friendly terms
User Experience: Making Complexity Simple
The biggest challenge? Making advanced mathematics accessible to busy marketing professionals.
I built a Python-based GUI using Tkinter that:
- Requires zero technical knowledge to operate
- Processes complex optimizations in 2-3 seconds
- Presents results in familiar marketing metrics (CPC, CPM, CTR)
- Includes built-in validation to prevent unrealistic scenarios
Real-World Impact and Applications
For Marketing Teams
- Faster decision-making: Get optimal budget allocation in seconds, not hours
- Better predictions: Know what to expect before launching campaigns
- Reduced uncertainty: Make data-driven decisions with confidence
- Cross-platform optimization: Manage Facebook, Google, and DV360 holistically
For the Industry
This research opens new possibilities for marketing technology:
- Real-time budget optimization during active campaigns
- Bi-objective optimization (maximize results while minimizing costs)
- Integration with existing marketing stacks and APIs
- Scalable solutions for agencies managing hundreds of clients
Technical Deep Dive: For the Data Science Community
Methodology Highlights
MINLP Implementation:
- Used Bonmin solver for mixed-integer nonlinear optimization
- Developed separate sub-models for conversion and awareness objectives
- Implemented diminishing returns through logarithmic utility functions
- Conducted sensitivity analysis with ±10% parameter variations
XGBoost Implementation:
- Hyperparameter optimization using Optuna (1000 trials)
- Feature engineering including platform-specific variables
- 5-fold cross-validation for robust performance evaluation
- Statistical significance testing against baseline models
System Architecture:
- Modular design enabling easy model updates
- Error handling and edge case management
- Performance monitoring and logging
- Scalable framework for additional platforms
Key Research Contributions
- First MINLP application in marketing mix modeling with empirical validation
- XGBoost effectiveness demonstration in non-time-series marketing contexts
- Proof-of-concept framework for OR-ML integration in marketing
- Practical implementation guide for industry adoption
Lessons Learned and Future Directions
What Worked Exceptionally Well
- MINLP’s robust performance across different scenarios and sensitivity tests
- XGBoost’s superior accuracy compared to traditional regression approaches
- User interface design that abstracts complexity without losing functionality
- Integration architecture that allows independent model improvements
Areas for Enhancement
- Larger dataset representation for high-budget campaigns
- Real-time data integration for dynamic model updating
- Multi-objective optimization to balance competing goals
- Extended platform support beyond the current three
Future Research Directions
- Bi-objective MINLP for cost minimization and result maximization
- Real-time optimization during active campaign periods
- Deep integration with marketing automation platforms
- Omnichannel campaign optimization across all touchpoints
The Technology Stack
Core Technologies:
- Python for model development and integration
- MINLP with Bonmin solver for optimization
- XGBoost with Optuna for machine learning
- Tkinter for GUI development
- NumPy/Pandas for data processing
- Scikit-learn for validation and metrics
Research Tools:
- Statistical testing for model validation
- Cross-validation for robust performance assessment
- Sensitivity analysis for model robustness
- Performance benchmarking against industry standards
Impact and Recognition
This research represents a significant advancement in marketing analytics, bridging the gap between academic research and practical industry applications. The system provides:
- Academic contribution: First integration of MINLP and XGBoost for marketing mix modeling
- Industry value: Practical framework that marketing teams can implement
- Performance validation: Rigorous testing with statistical significance
- Future foundation: Extensible architecture for continued development
Ready to Discuss This Project?
I’m passionate about applying advanced analytics to solve real business problems. This thesis project demonstrates my ability to:
✅ Tackle complex, unsolved problems with innovative approaches
✅ Bridge academic research and industry needs effectively
✅ Develop end-to-end solutions from concept to working prototype
✅ Validate results rigorously with proper statistical methods
✅ Communicate technical concepts to diverse audiences
Want to see this system in action or discuss how similar approaches could benefit your organization?
📧 Email me for a discussion
📱 Connect with me on LinkedIn
📄 Download the full thesis document (soon)
💻 View the code repository on GitHub
For a deeper dive into the methodologies and data, the project is available in its GitHub repository and will soon be released on IU University of Applied Science’s archive.
Keywords: marketing mix modeling, MINLP, XGBoost, machine learning, operations research, marketing analytics, campaign optimization, predictive analytics, data science