aave-wallet-segmentation

A wallet segmentation analysis provided for Aave Grants DAO

View readme in google docs at https://docs.google.com/document/d/1qPe5P7X9FhsVrNgdny8CuE81-3dOJ7LVAOKCvPnsn5M/edit?usp=sharing

Aave Grant Final Submission

Aave Wallet Segmentation Profiles

• Final Submission Video Presentation
• Final Presentation Slides Doc PDF
• Final Write Up Doc PDF

Intro / TL;DR

The Fraud Detection & Defense workstream at Gitcoin applied and was approved for a grant to perform a wallet segmentation analysis for the Aave ecosystem. Using their unique understanding of onchain behaviors and analysis provided by 2 years defending Gitcoin from Sybil attacks, they put their team to work understanding the user groups & personas which interact with Aave contracts.

• The analysis clearly identifies 13 user personas
• The 13 personas are bucketed into 3 categories: Testers, Income, & Special Cases

Recommendations for insights

• Improve targeting & copy for marketing & to focus on revenue generating users
• Use to qualify participants in product research
• Consider custom experiences for highest value user personas*

Potential to build on this work

• Open source work can be built on by the community
• Consider time based behavioral analysis and retention analysis
• Create a system for identifying how this analysis is used to measure impact
• Run a targeted campaign to incentivize qualitative feedback at scale

Data Discovery & Cleaning

• Previous Aave grant had already documented all data sources
• Contract calls: deposit, supply, borrow, repay, withdraw, redeemunderlying, liquidationcall, flashloan
• Chains: Arbitrum, Avalanche, Polygon, Fantom, Optimism, Ethereum
• Versions: 1,2,3
• Exogenous data sources: Lens, Snapshot, Trustalabs, Gitcoin, Debank
• Removed outliers, reduced sample size, merging of function calls
• Profiling based on 114,915 wallets available on mainnet for data availability
• Histograms: by chain, by event, by version
• Scatterplot matrix of all variables endogenous & exogenous

Methodology

• Non-linear dimensional reduction: t-SNE & UMAP
• Manual parameter search looking for meaningful separation
• Visual investigation of 3-Dimensional feature space
• Plotting multiple 2 dimensional projections of UMAP space
• Linking graphs with an interactive table to track cluster averages
• Manual brushing to investigate pairs of clusters for cohesion, compactness, & quantitative distinctness
• Reviewing the final cluster solution in 2D & 3D space

Results (Personas)

• Quantitative review of the group mean vectors using color g* radient for examination
• Parallel coordinate plots to visualize segments in multiple dimensions
• Personas created based on behavioral observations:

Special Cases

• The Good Guys (18,512 | 16.11%)
• The Liquidated (2,324 | 2.02%)
• The Throwaway Accounts (3,191 | 2.78%)
• The Potential Arbitragers (3,824 | 3.33%)

Income

High Rollers

• Without debt (6,553 | 5.7%)
• With debt (7,016 | 6.11%)

Middle Class

• High checking, low savings (10,174 | 8.85%)
• High savings, low checking (9,477 | 8.25%)

Small Frys

• Depositors on Ethereum (2,773 | 2.41%)
• Degen Active Depositors (1,202 | 1.05%)
• Debt Users (16,529 | 14.38%)

Testers

• Ethereum Only (9,321 | 8.03%)
• Multichain (24,107 | 20.98%)

• Review table of proportions, counts, and variables against exogenous variables