---

📌 Project Overview

E-commerce businesses lose millions every year to cart abandonment and missed conversions. This project builds a binary classification model that predicts whether an online shopping session will end in a purchase — using behavioral signals like page views, bounce rates, exit rates, and session metadata.

The goal: give e-commerce teams a reliable signal to personalize UX or trigger real-time interventions for high-intent visitors.

• Task: Binary Classification (Revenue: Purchase or No Purchase)
• Dataset: Online Shoppers Purchasing Intention Dataset
• Best Model: Decision Tree Classifier (Tuned via GridSearchCV)

---

📂 Dataset

Property	Value
Total Samples	12,329
Features	17 (10 numerical, 7 categorical)
Target	Revenue (True/False → 1/0)
Class Imbalance	~84.5% No Purchase / ~15.5% Purchase
Missing Values	None

Key features:

• PageValues — average value of pages visited before transaction
• ExitRates — average exit rate of pages visited
• BounceRates — average bounce rate of pages visited
• ProductRelated, ProductRelated_Duration — product page engagement
• Month, VisitorType, Weekend — session context

---

🔄 Pipeline Workflow

Raw CSV → EDA → Feature Split → Train/Test Split → Preprocessing Pipeline → Model Training → GridSearchCV → Evaluation

1️⃣ Load & Inspect — Load CSV, check nulls, understand class distribution
2️⃣ Feature Separation — Split numerical vs categorical columns automatically using select_dtypes
3️⃣ Train/Test Split — 80/20 split with stratify=y to preserve class ratios
4️⃣ Preprocessing Pipeline — StandardScaler for numerics, OneHotEncoder for categoricals via ColumnTransformer
5️⃣ Model Training — DecisionTreeClassifier with class_weight="balanced" to handle imbalance
6️⃣ Hyperparameter Tuning — GridSearchCV over max_depth and min_samples_leaf
7️⃣ Evaluation — F1 Score, Precision, Recall, Classification Report, Confusion Matrix

---

🤖 Models

1️⃣ Decision Tree Classifier ⭐ Best Model

PrePruning:

dtc_model = DecisionTreeClassifier(
    max_depth=6,
    min_samples_leaf=30,
    class_weight="balanced",
    random_state=42
)

pipe = Pipeline(steps=[
    ("preprocess", preprocessor),
    ("model", dtc_model)
])

• class_weight="balanced" corrects for the 84/16 class imbalance automatically
• max_depth=6 prevents overfitting on noisy behavioral signals
• min_samples_leaf=30 ensures stable leaf nodes with sufficient support
• Best GridSearchCV params: max_depth=4, min_samples_leaf=50

---

📊 Results

Model	Precision	Recall	F1 Score	Accuracy
Decision Tree (Baseline)	0.504	0.833	0.628	84.71%
🏆 Decision Tree (Tuned)	0.498	0.833	0.623	84.39%

Note: High recall (83.3%) is the key optimization target — missing a real buyer (False Negative) costs more than flagging a non-buyer (False Positive). The class_weight="balanced" parameter drives this recall-first behavior.

---

🔍 Key Insights

• 📈 PageValues has a +0.49 positive correlation with Revenue — the single strongest predictor; users who browse high-value pages are far more likely to purchase
• 📉 ExitRates has a -0.21 negative correlation with Revenue — high exit rates are a clear signal of abandonment intent
• 📉 BounceRates has a -0.15 negative correlation with Revenue — single-page sessions rarely convert
• ProductRelated page engagement (both count and duration) correlates positively at +0.15, confirming deeper product browsing = higher purchase probability
• SpecialDay proximity slightly reduces purchase probability (-0.08) — possibly due to browsing-only behavior during holidays
• class_weight="balanced" was critical: without it, the model predicts mostly non-purchase due to 84/16 class split

---

🗂️ Repository Structure

shop-smart-ecommerce/
│
├── shop_smart_ecommerce.ipynb   # Main notebook: EDA, training, evaluation
├── shop_smart_ecommerce.csv     # Dataset (Online Shoppers Purchasing Intention)
└── README.md                    # Project documentation

---

🚀 Quick Start

# Clone the repository
git clone https://github.com/ronakrajput8882/Shop-Smart-Ecommerce.git
cd Shop-Smart-Ecommerce

# Install dependencies
pip install pandas scikit-learn seaborn matplotlib jupyter

# Launch notebook
jupyter notebook shop_smart_ecommerce.ipynb

---

🧠 Key Learnings

• Imbalanced classification requires intentional design — class_weight="balanced" shifted the model from 0% recall on the minority class to 83%, without any oversampling
• Sklearn Pipelines prevent data leakage — fitting the scaler only on train data, applied on test, is the correct production pattern
• High recall ≠ high precision in imbalanced tasks — depending on business cost, optimizing for recall over F1 or accuracy may be the right call
• PageValues dominates — a single engineered feature derived from Google Analytics carries more signal than all session-time features combined
• GridSearchCV with scoring="f1" ensures tuning aligns with actual business metric, not just accuracy

---

🛠️ Tech Stack

Tool	Use
Python 3.10+	Core language
Pandas	Data loading & manipulation
Scikit-learn	Pipeline, preprocessing, model, GridSearch
Seaborn / Matplotlib	EDA visualization
Jupyter Notebook	Interactive development

---

Connect with me

If you found this useful, please ⭐ the repo!!!