← Back to Home
NETWORK VISUALIZATION • CASE STUDY
User Journey & Traffic Flow Analysis via Plotly Sankey
1. Dataset, Scope & Business Context
Project Type
Personal Project / Simulated Business Case (Maven Fuzzy Factory)
Dataset Source
Maven Fuzzy Factory Teaching Dataset (5 CSVs: orders, order_items, products, website_sessions, website_pageviews)
Data Scale
32,313 orders • 472,871 sessions • 1,188,124 pageviews
Time Range & Currency
Mar 2012 – Mar 2015 (3 Years) • USD ($)
Standard aggregate bar charts failed to capture how traffic sources filtered into landing pages or the exact product transitions adopted by repeat buyers across multi-touch sessions.
- Build an Acquisition Sankey Flow:
UTM Source → Device → Landing Page → Conversion / Drop-off. - Build a Retention Sankey Flow:
1st Purchase → 2nd Purchase → 3rd Purchase. - Apply visual data storytelling via selective opacity to emphasize "happy paths" (Conversions & Repurchases) while muting terminal drop-offs.
2. Key Findings & Behavioral Flow Insights
- Denominator Scope & the "Engaged Leaver" Pattern: Out of 472,871 total sessions, 211,640 (44.8%) were single-pageview bounce sessions and were excluded via a
pageviews > 1filter so Sankey 1 focuses strictly on the 261,231 (55.2%) engaged multi-page sessions. Even within this engaged denominator, the mid-funnel drop-off node dominates overConverted—pinpointing catalog/checkout UX hypotheses beyond initial ad-click quality. - Device Conversion Discrepancy: Desktop sessions flow smoothly into the
Convertednode, whereas Mobile traffic disproportionately terminates at theDrop-offnode. - The Gateway & Cross-Pollination Journey:
- The Original Mr. Fuzzy serves as the primary top-of-funnel gateway item, dominating Node 1.
- Repeat-purchase paths consistently transition from Mr. Fuzzy (Purchase 1) → Love Bear (Purchase 2). Customers rarely purchase the exact same plush toy twice, favoring cross-catalog exploration.
3. Methodology & Analysis Fixes
- Preventing Circular Link Kernel Crashes: Plotly Sankey diagrams fail when target nodes loop back into identical source nodes. Resolved by prefixing chronological order rankings (
1 - Product A→2 - Product A), ensuring every node in the directed acyclic graph (DAG) remains strictly unique. - Memory Optimization on 1M+ Row Joins: Reduced RAM overhead by specifying
usecolsduringpd.read_csv()and aggregating standalone tables via.groupby().first()before performing final joins. - Visual De-Noising via Dynamic Link Alpha: Engineered an automated link-coloring loop: assigned terminal paths (Drop-off/End) a muted gray
rgba(200, 200, 200, 0.4)while highlighting converted progression paths in vivid amber/emerald. - Standardized Acquisition Imputation: Cleaned fragmented
NaNand arbitraryunknownUTM source strings into a unified, business-readableDirect Trafficnode.
4. Proposed UX & CRM Experiments
- Mobile UI/UX Audit: Test sticky "Add to Cart" CTAs and lighter mobile media payloads to test whether mobile checkout friction drives the observed drop-off discrepancy.
- Lifecycle Cross-Sell Trigger: Suppress Mr. Fuzzy promos to existing Mr. Fuzzy owners and A/B test a 14-day post-purchase Love Bear email trigger aligned with the observed repurchase flow.
Evaluation Metrics: Track Mobile Cart-to-Order Conversion Rate (UX audit) and Incremental 30-Day Second-Purchase Rate vs. control group (CRM trigger).