PythonScikit-learnLogistic RegressionRandom ForestEDA

Conversion Rate Optimization

ML model that targets the top 10% of users responsible for 95% of conversions

Overview

Analyzed e-commerce user behavior data across 300K+ sessions to identify the drivers of conversion and built a classification pipeline that ranks users by purchase probability. The model enables targeted marketing campaigns with measurable ROI — at a $1 outreach cost and $40 profit per conversion, only a 2.5% predicted probability is needed to break even.

Methods

EDA across traffic source (Ads, SEO, Direct), country, age, and session engagement segments

Logistic Regression with one-hot encoding of categorical features

Random Forest (200 estimators) as a nonlinear comparison model

Lift curve and decile analysis to quantify campaign targeting efficiency

Feature importance via logistic coefficients and Random Forest impurity scores

Secondary model excluding session engagement to isolate static user signals

Key Findings

Session engagement (pages visited) accounts for ~90% of Random Forest feature importance — the single most predictive behavioral signal
Returning users convert at 7.2% vs. 1.4% for new users — a 5× difference
Top 10% of users scored by the model capture 95% of all conversions
Germany and UK users convert at 6.3% and 5.3%, nearly double the US rate
Without session engagement, model AUC drops from 0.986 to 0.82 — confirming it as the dominant predictor
The decision problem is near-linearly separable in log-odds space, explaining why Logistic Regression outperforms Random Forest

Results

ROC AUC: 0.986 (Logistic Regression), 0.975 (Random Forest)

Precision 0.85, Recall 0.69 on converters at default threshold

Top decile conversion rate: 30.3% vs. 3.2% overall average — a 9.5× lift

Targeted campaigns viable at ≥2.5% predicted probability, far below the model's score range for high-confidence users

Ask About This Project

Have a technical question? Ask here.

← All projectsEric Jin