The Algorithm

We measure social distance between borrower and lender using the Adamic-Adar Index—a proven network science algorithm that weights selective connections higher.

---

The 4-Step Process

1. Find Mutual Connections

B = borrower's network (followers + following)
L = lender's network (followers + following)
M = mutual connections (B ∩ L)

---

2. Adamic-Adar Weighting

Core insight: A mutual friend with 20 connections is a stronger signal than one with 20,000.

For each mutual connection:
  weight = 1 / log(total_connections)

AA_score = Σ weights for all mutuals

Example:

• Friend with 25 connections → weight = 0.31

• Friend with 10,000 connections → weight = 0.11

• Small network weighted 3× higher

The logarithm provides diminishing returns as networks grow, preventing influencer gaming while rewarding tight-knit communities.

---

3. Quality Adjustment

Filter spam/bots using Neynar quality scores:

Q_avg = (Q_borrower + Q_lender) / 2
AA_effective = AA_score × Q_avg

Quality scores (0-1 scale):

• Legitimate user: 0.9

• Low activity: 0.5

• Bot/spam: 0.1

---

4. Calculate Final Score (0-100)

Three components (initial thresholds):

A. Base Score (max 60 points)

AA_effective	Points
≥ 20	60
≥ 10	50
≥ 5	35
≥ 2.5	20
≥ 1	10

B. Overlap Bonus (max 30 points)

When significant portion of both networks overlap:

P_overlap = (M / min(|B|, |L|)) × 100
Bonus = min(P_overlap × 3, 30) if P_overlap > 10%

C. Mutual Follow Bonus (max 10 points)

• Both follow each other: +10

• One-way follow: +5

• No direct follow: +0

Final:

S_total = min(S_base + S_overlap + S_mutual, 100)

Ongoing Refinement

The specific thresholds (≥20, ≥10, etc.), bonus multipliers, and score caps are initial parameters based on network science research. We'll continuously refine these as we collect repayment data to optimize prediction accuracy.

---

Example

Given:

• 25 mutual connections

• AA_score = 8.5

• Q_borrower = 0.9, Q_lender = 0.85

• Both follow each other

Result:

1. AA_effective = 8.5 × 0.875 = 7.44

2. Base = 35 points

3. Overlap = 0 (below 10% threshold)

4. Mutual follow = 10 points

5. Total = 45 (MEDIUM RISK)

---

Why This Works

Research shows:

• Adamic-Adar improves link prediction accuracy by 82% over simple friend counting [7]

• Social network ties on Prosper.com correlate with better repayment performance [8]

• Consistently top-3 performing algorithm for network similarity measurement

Note: The algorithm was designed for link prediction, not loan defaults. We're testing whether it translates to repayment prediction in Phase 0.

Benefits:

• Friends know you better than strangers (reduces information asymmetry)

• Penalizes influencer gaming (following 10K people = minimal signal)

• Rewards tight communities (close-knit groups score highest)

• Sybil resistant (quality filtering + weighting)

• Transparent (public social graph data)

---

Back to: Social Trust Scoring · Risk Scoring