Business Analytics • Conversion Modeling

Customer Purchase Prediction & Conversion Analytics

Predicting purchase likelihood from browsing behavior, visitor context, model evaluation, and threshold-tuned conversion targeting.

This project builds a purchase-propensity workflow for online shopping sessions. It combines exploratory conversion analysis, supervised classification, threshold optimization, and feature-importance interpretation to support business-focused targeting decisions.

Gradient Boosting • Logistic Regression • Random Forest • Threshold Tuning • Feature Importance

12,205 Cleaned shopping sessions

15.63% Observed conversion rate

0.9377 Gradient Boosting ROC-AUC

0.70 Selected decision threshold

Project Overview

This project analyzes online shopping sessions to predict purchase likelihood and support conversion targeting. The workflow combines exploratory customer behavior analysis, class-imbalance-aware modeling, model comparison, threshold tuning, and feature-importance interpretation to connect predictive performance with practical business decision-making.

Prediction Built classification models to predict purchase outcomes from browsing behavior, product engagement, traffic context, and visitor attributes.

Conversion Analyzed purchase imbalance, visitor segments, behavioral differences, and dominant conversion signals such as PageValues and product engagement.

Targeting Selected a Gradient Boosting model and tuned the purchase-probability threshold to balance recall, precision, and campaign actionability.

Conversion Behavior

The dataset is highly business-oriented: Most online shopping sessions in the dataset do not result in a purchase, creating an imbalanced conversion problem. Before modeling, the project analyzes purchase distribution, visitor-type conversion rates, and behavioral value signals to understand which session patterns are associated with revenue.

Purchase outcome distribution showing purchase and non-purchase sessions — **Purchase outcome distribution.** The dataset contains many more non-purchase sessions than purchase sessions, making recall and precision more informative than accuracy alone.

Conversion rate by visitor type — **Visitor-type conversion.** Conversion rates differ across visitor categories, showing that user context carries meaningful business signal.

Log-scaled PageValues distribution by purchase outcome — **PageValues signal.** PageValues strongly separates purchase and non-purchase sessions, making it one of the most important behavioral indicators in the conversion workflow.

Modeling Workflow

The notebook compares multiple supervised classification models and evaluates them using business-relevant metrics. The workflow progresses from cleaned session data and preprocessing through model comparison, threshold tuning, and final business-oriented evaluation of purchase-likelihood predictions.

Model Performance

Gradient Boosting was selected as the final model because it delivered strong ranking performance and supported threshold-based targeting. At the selected threshold of 0.70, the model identifies a substantial share of purchase sessions while limiting unnecessary targeting of non-purchase sessions.

Final model	Accuracy	Precision	Recall	F1-score	ROC-AUC
Gradient Boosting at threshold 0.70	89.59%	64.35%	75.13%	0.6932	0.9377

Gradient Boosting confusion matrix at threshold 0.70 — **Final confusion matrix.** Final confusion matrix. At threshold 0.70, the model correctly identified 287 of 382 purchase sessions in the test set, while missing 95 purchase sessions and incorrectly flagging 159 non-purchase sessions.

Business Interpretation

The selected threshold balances identifying likely buyers with reducing unnecessary targeting of non-purchase sessions. In a business setting, this type of model could support campaign targeting, remarketing prioritization, or conversion-likelihood scoring rather than replacing human strategy.

Threshold Optimization

Instead of relying on the default 0.50 classification threshold, the project evaluates multiple probability cutoffs. This is important because conversion modeling is usually a business tradeoff: a lower threshold captures more buyers but may target more non-buyers, while a higher threshold improves confidence but may miss opportunities.

Gradient Boosting threshold tradeoff curve showing precision, recall, and F1-score — **Threshold tradeoff.** Precision, recall, and F1-score were evaluated across probability thresholds. The 0.70 cutoff was selected because it improved prediction reliability while still capturing a substantial share of purchase sessions.

Interpretation

The threshold analysis turns the model from a generic classifier into a business decision tool. A threshold of 0.70 provides the strongest practical balance between identifying purchase-likely sessions and reducing unnecessary targeting of non-purchase sessions.

Feature Signals and Interpretation

The model interpretation step identifies the behavioral signals that contribute to purchase-likelihood scoring. PageValues was the dominant predictor in the Gradient Boosting model, so this secondary feature-importance view removes it from the chart to reveal additional predictors beyond the strongest single signal.

Secondary Gradient Boosting feature importance without PageValues — **Secondary feature importance.** After excluding the dominant PageValues feature from the visualization, the next strongest signals include Month: November, product-related duration, exit rate, and browsing-depth features.

Role and Implementation

I implemented the full conversion analytics workflow, including data cleaning, exploratory analysis, leakage-safe preprocessing, model comparison, threshold tuning, final model evaluation, and feature-importance interpretation.

Conversion Analysis

Analyzed purchase imbalance, visitor behavior, PageValues dominance, and session-level engagement patterns before modeling.

Predictive Modeling

Compared Logistic Regression, Random Forest, and Gradient Boosting models using a consistent preprocessing pipeline.

Threshold and Evaluation

Tuned the final Gradient Boosting decision threshold and evaluated performance using precision, recall, F1-score, ROC-AUC, and confusion matrices.

Tools

Python • Pandas • NumPy • Matplotlib • Scikit-learn • Seaborn • Jupyter Notebook

Artifacts

The project artifacts include the notebook, source repository, conversion analysis visuals, model evaluation outputs, and interpretation graphics.

GitHub Repository

Source repository containing the notebook, code, visual outputs, and documentation.

Open repository →

Notebook

Main Jupyter Notebook covering conversion analysis, model comparison, threshold tuning, and feature interpretation.

customer_purchase_prediction_conversion_analytics.ipynb

Visual Outputs

Selected visuals for conversion behavior, feature signal, threshold optimization, final model performance, and secondary feature importance.

images/projects/customer-purchase/