A machine-learning–powered simulation of how media companies like NBCUniversal or Warner Bros. Discovery could optimize cross-channel advertising budgets.
OnePlan AI models how different media channels (TV, Streaming, YouTube, Display, Social) respond to ad spend and uses machine learning + mathematical optimization to recommend the most efficient cross-channel allocation.
It replicates the principles behind NBCUniversal’s One Platform Total Audience system — bringing together data science, optimization, and explainability.
Media planners struggle to understand where each marketing dollar delivers the most conversions or reach.
This project answers one key question:
“Where should the next dollar go to maximize ROI?”
| Layer | Description | Notebook |
|---|---|---|
| 1. Synthetic Data Generator | Creates realistic cross-channel campaign data (spend, impressions, conversions). | 01_synthetic_data_generator.ipynb |
| 2. EDA & Media Math | Validates channel behavior, CPM logic, and correlations between spend & performance. | 02_eda_and_media_math.ipynb |
| 3. ML Modeling (Linear & XGBoost) | Fits non-linear response curves with Savitzky–Golay smoothing to reduce noise. | 03_model_response_curves.ipynb |
| 4. Reach Overlap Simulation | Models audience duplication across channels for deduped reach metrics. | 04_reach_overlap_simulation.ipynb |
| 5. Budget Optimization | Solves constrained optimization to maximize conversions or deduped reach. | 05_budget_optimizer.ipynb |
| 6. Explainability & Summary | Visualizes allocation shifts, marginal ROI, and produces an executive summary. | 06_explainability_and_summary.ipynb |
File:
assets/response_curves_overlay.png
Shows spend → predicted conversions per channel.
- Steeper slopes → more efficient channels (YouTube, Social).
- Flattened curves → saturation (Linear TV).
File:
assets/p6_opt_spend_conversions.png
Bar chart of budget by channel for the conversions objective.
File:
assets/p6_marginal_roi.png
Shows the extra conversions per $1 at optimized allocation.
High marginal ROI → channel still has growth potential.
Files:
assets/p6_equal_vs_opt_conversions.pngassets/p6_equal_vs_opt_reach.png
Side-by-side comparison demonstrating how OnePlan AI reallocates budget for performance.
| Metric | Equal Split | Optimized | Lift % |
|---|---|---|---|
| Conversions | 111.9 | 121.4 | +8.5 % |
| Deduped Reach | 61.1 % | 45.8 % | −25.0 % |
The optimizer increased total conversions by +8.5 % by shifting spend toward high-ROI digital channels, even though deduplicated reach decreased by about 25 % — a realistic efficiency vs reach trade-off.
- Executive Summary:
reports/OnePlan_Executive_Summary.md
Non-technical overview explaining what changed, why the optimizer reallocated spend, and how much lift it produced. - Model Performance:
data/processed/model_response_summary.csv
RMSE / MAE / R² per channel (Linear Regression vs XGBoost).
- Digital channels (Social & YouTube) deliver higher marginal ROI at lower CPMs.
- Linear TV saturates quickly — still valuable for reach but inefficient for conversions.
- Optimizer reallocates ~20 % of budget to digital → ~8–10 % conversion gain.
- Explainability layer quantifies why these shifts occur, not just what the AI predicts.
Python 3.11 · Pandas · Scikit-learn · XGBoost · SciPy · Matplotlib
Developed and tested on macOS using Jupyter (venv).
notebooks/Jupyter analysis parts + verify notebook data/raw/ synthetic data + overlap matrix data/processed/response curves, optimizer outputs, metrics assets/plots used in README and reports reports/ OnePlan_Executive_Summary.md (+ optional HTML) README.md this file LICENSE (added in next step) requirements.txt pinned dependencies (optional)
This project was independently created for educational and portfolio demonstration purposes. Inspirations:
- NBCUniversal One Platform Total Audience (media planning & reach analytics)
- Nielsen & Comscore public documentation on audience modeling
- SciPy, Scikit-learn, and XGBoost open-source projects
2025 Aravind Anisetti. All rights reserved.