Project Overview

In an increasingly competitive digital marketplace, understanding what drives users to make purchases is critical for optimizing websites and increasing conversion rates. This project investigates the browsing behavior of users and how various on-page actions — viewing images, reading reviews, checking product specifications, comparing similar items — influence their purchasing decisions.

The primary output is a deployed lead scoring model that classifies sessions by purchase likelihood, allowing sales and marketing teams to focus intervention efforts on the highest-intent prospects rather than treating every visitor identically.

Data Dictionary

The dataset captures binary interaction signals from 500 user sessions on an e-commerce product page. Each variable records whether a user engaged with a specific page element:

Executive Summary

Out of 500 user sessions, 185 resulted in a purchase (37%) and 315 did not (63%). This class imbalance was a key challenge addressed in the modeling phase. The analysis identifies which browsing behaviors are most predictive of purchase intent — and which combinations create the strongest synergy.

Exploratory Data Analysis

Class Distribution

The target variable showed a meaningful class imbalance — a critical finding that shaped the modeling approach:

BUY
0    315    (63.0%) — did not purchase
1    185    (37.0%) — made a purchase
Total: 500 sessions

With a 63/37 split, naive models trained on raw data would be biased toward predicting non-purchases. Four resampling techniques were evaluated to address this: Random Under-Sampling, Random Over-Sampling, SMOTE, and NearMiss.

Analytical Techniques Applied

• Summary Statistics — mean, median, variance, and standard deviation for all binary interaction variables
• Contingency Tables — crosstab analysis between each behavioral variable and the BUY outcome to measure individual impact
• Correlation Heatmap — visualizing which features had the strongest linear relationship with purchase behavior
• Interaction Heatmaps — pairwise analysis of feature combinations to identify synergistic effects on conversion

Interaction Analysis

The most valuable EDA finding was how combinations of page elements affect purchase likelihood — not just individual elements. Each interaction pair was analyzed as a 2×2 contingency table.

Key Takeaways from Interaction Analysis

Machine Learning Models

Four resampling techniques were evaluated to address the 63/37 class imbalance before training a classification model. Each approach handles imbalance differently — the comparison reveals which produces the most reliable predictions:

Model Performance Comparison

All four models were evaluated on precision, recall, F1-score, and accuracy. Random Over-Sampling and SMOTE both achieved 99% accuracy with near-perfect precision and recall across both classes:

Detailed Results: Best Models

              precision    recall  f1-score   support

           0       1.00      0.98      0.99       100
           1       0.98      1.00      0.99        89

    accuracy                           0.99       189
   macro avg       0.99      0.99      0.99       189
weighted avg       0.99      0.99      0.99       189

              precision    recall  f1-score   support

           0       1.00      0.98      0.99       100
           1       0.98      1.00      0.99        89

    accuracy                           0.99       189
   macro avg       0.99      0.99      0.99       189
weighted avg       0.99      0.99      0.99       189

Live Demo

The lead scoring model is deployed as an interactive Streamlit application. Input a set of browsing behaviors and get an instant purchase likelihood score:

Recommendations

Highlight Product Reviews and Images Together

• Ensure reviews are prominently displayed alongside product images — not buried in a separate tab. The 58.33% conversion rate from this combination is the single most actionable finding in this study.
• Use verified purchase reviews rather than unverified ratings — the signal is stronger when users trust the source of the review.
• Display the review count prominently: social proof scales with volume. More reviews shown = higher perceived credibility.

Engage Users with Content That Reduces Uncertainty

• Pair product specifications with prominent warranty information — this combination crosses the 50% purchase threshold (51.85%) and removes two categories of buyer objection simultaneously.
• Consider adding customer testimonial videos or comparison charts — active engagement with product information is strongly predictive of purchase intent.
• Ensure FAQ content is paired with visuals — FAQ alone converts at only 21.74%, but with images this rises to 42.86%.

Optimize for Comparison Behavior

• Users who compare similar items convert at 33.33% — higher than sponsored link clickers (23.81%). Design the comparison UI to be smooth and fast, as this behavior signals strong purchase intent.
• Consider triggering a targeted offer or chat prompt when a user initiates a product comparison — this is a high-intent behavioral signal.

Deploy the Lead Scoring Model

• Use the trained model to score live sessions in real time, enabling sales teams to prioritize outreach to high-intent visitors.
• Build a trigger system: sessions with predicted purchase probability above a threshold (e.g., 70%) receive a personalized intervention — live chat offer, limited-time discount, or follow-up email.

Limitations & Further Research

Data Limitations

• The dataset captures only 500 sessions with a limited set of on-page interactions — broader behavioral signals (scroll depth, time-on-page, mouse movement) could significantly improve predictive power.
• No pricing, promotional, or competitor context is captured. Purchase behavior is heavily influenced by price sensitivity and competitive availability, which are absent from this model.
• No demographic or device information — mobile vs. desktop behavior, age, and location are known moderators of purchase intent that are unaccounted for.

Further Research

• Temporal analysis — explore how conversion patterns change during sales events, holiday seasons, or promotional periods.
• User segmentation — segment users by demographics or acquisition channel and build segment-specific models; a first-time visitor from paid ads may require different signals than a return visitor from organic search.
• Sequential behavioral modeling — the order in which users engage with page elements (e.g., images first vs. reviews first) may contain predictive signal not captured in the current binary features.

Related Projects

If this case study is relevant to your business challenge, these projects may also interest you: