Click-Through-Rate (CTR) Prediction  —  Classical Machine-Learning Pipeline

Goal: build and evaluate a lightweight yet accurate model that predicts
the probability a user will click an online advertisement, using only classical (non-deep) ML techniques and modest hardware.

---

Project Structure

.
├── data/                  # raw & processed CSV files
├── notebooks/
│   └── CTR_EDA_and_Model.ipynb
├── plots/                 # 9 key figures exported from the notebook
│   ├── output.png
│   ├── output2.png
│   └── … output9.png
├── src/
│   ├── features.py        # feature-engineering pipeline
│   ├── train.py           # model training / cross-validation
│   └── utils.py
├── environment.yml        # Conda environment (Python 3.11)
├── CTR_Report.tex         # full 14-page LaTeX report
└── README.md              # <── you are here

---

⚙️ Setup

# clone repository
git clone https://github.com/Mah-En/Click-Through-Rate-Prediction-in-Online-Advertising
cd CTR_Project1

# create environment
conda env create -f environment.yml
conda activate ctr

# (optional) install as editable pkg
pip install -e .

Main dependencies ≈ scikit-learn 1.5, pandas 2, matplotlib 3.9, xgboost 2.

---

Quick Start

# 1. run end-to-end pipeline (train, evaluate, export model.pkl)
python src/train.py --config configs/gbdt.yaml

# 2. reproduce every plot + PDF report
make all                       # ⇢ runs notebook, saves figures, builds LaTeX

After make all you’ll find:

• plots/output*.png – nine explanatory figures
• CTR_Report.pdf – 14-page article-style report ready to submit

---

Key Findings

Model	AUC-ROC	Log-Loss	Brier	Cost ($)
Logistic Regression	0.925	0.211	0.138	0.643
Random Forest	0.964	0.146	0.097	0.522
Gradient Boosting	0.974	0.121	0.084	0.505
XGBoost	0.973	0.124	0.086	0.511

• Tuned GBDT cuts business loss by 21 % vs. a naïve benchmark
• Median inference latency 4 ms on a Raspberry Pi 4, model size 1.9 MiB
• Fairness gaps (gender & age) < 0.03 ⇒ passes 5 % parity rule
• Feature importance: Engage Index (Time / Age), Hour-of-Day, encoded City

---

📊 Exploratory Plots

Histogram bundle	Gender counts	Age vs Click

(see /plots for the full set)

---

Reproducing the Report

1. Ensure pdflatex is in your path (TeX Live 2025 or MiKTeX).

2. Run make report or

   pdflatex CTR_Report.tex
   pdflatex CTR_Report.tex   # run twice for cross-refs

No BibTeX required; references are hard-coded.

---

Next Steps

• Cost-sensitive training (loss weights during tree growth)
• LightGBM for ≥ 1 M impressions
• Enrich Ad Topic Line with TF-IDF or fastText embeddings
• Adversarial regularisation for counter-factual fairness

---

References

1. Interactive Advertising Bureau, Internet Advertising Revenue Report, 2024.
2. Richardson et al., “Predicting CTR for New Ads,” WWW 2007.
3. He et al., “Practical Lessons from Predicting Clicks on Ads at Facebook,” ADKDD 2014.

---