Network Strength Algorithm Development

The Advanced Problem: Network Scores Collapse Past 500 Nodes

Every practitioner who has shipped a network strength algorithm has hit the same wall. The first version ranks users beautifully on a 40-node test graph, and by the time the ego network crosses 500 nodes the score is dominated by noise. The failure is structural, not statistical. Three mechanisms cause it.

Degree inflation. Connection count is the cheapest signal to collect and the least informative to use. The distribution of accepted connection requests on most platforms is near-Poisson with a long right tail, which means the top decile of degree holders is populated mostly by recruiters, community managers, and people who accept everything. Weight by degree and you are effectively weighting by willingness to tap accept. Freeman's betweenness and closeness formulations were never designed to carry that load alone -- they describe position, not relationship quality.

Static snapshots. A graph exported once is a corpse. Professional ties have half-lives. A colleague you shipped a project with 18 months ago and have not spoken to since carries a materially different weight than one you messaged last week, yet both contribute 1.0 to an unweighted adjacency matrix. Workings.me treats temporal decay as a first-class design constraint rather than a post-hoc multiplier applied at the end of the pipeline.

Redundancy blindness. Ten ties inside the same team, same company, same city, and same chat workspace are close to one tie in information terms. Burt's structural hole research established that non-redundant contacts -- those reaching separate clusters -- explain more variance in access than total contact volume. Any aggregator that sums edge weights without a redundancy penalty systematically over-scores the densest, most insular networks and under-scores the sparsely but strategically connected.

Layer on survivorship bias in labeling (you only observe opportunities that converted) and monotemporal bias in sourcing (one employer, one conference, one bootcamp cohort), and you have the reason most network scores plateau below an r-squared of 0.2 against real outcomes.

The fix is not a better scalar. It is a layered model: a stratified edge-weight function, an eigenvector aggregation with damping, and an explicit redundancy correction. That is what the rest of this article builds. Practitioners who want a fast directional read on their own graph before instrumenting anything can run the Career Pulse Score first and use the output as a sanity baseline.

The Stratified Edge Model (SEM-7): A Named Framework for Composite Network Strength

SEM-7 starts from one premise: an edge is a vector, not a scalar. Seven signal classes define it. Five enter a convex weighted sum; two act as multipliers so they can suppress a relationship rather than merely dilute it.

Recency carries the largest single weight because professional ties decay faster than most modelers assume. Intent carries the second largest because a single shipped artifact, merged pull request, signed contract, or completed introduction outweighs forty reaction-emoji exchanges. Frequency is deliberately log-scaled: the information gain from the 2nd interaction to the 10th is large, and from the 200th to the 210th is approximately zero.

Channel diversity is a proxy for tie multiplexity. Granovetter's original work on the strength of weak ties showed that ties embedded in multiple overlapping contexts are more durable and more likely to transmit novel information than single-context ties. Operationally, a contact you share a repository with, have met at a conference, and consult with quarterly scores materially higher on D than one you only ever email.

The two multipliers exist for a reason. Structural position should amplify or suppress, not add -- a tie to a genuine bridge is worth multiples of a tie to the eighth person inside your current employer. Identity confidence must be able to zero out an edge entirely when the record is a stale duplicate or an unverified alias. Workings.me models both as multipliers for exactly this reason, and surfaces the resulting single index alongside the underlying decomposition so users can see which component moved.

A practical calibration note: fit the five weights against your own outcome variable rather than adopting defaults wholesale. A 90-day logistic regression with elastic net regularization on labeled conversions (introductions that led to a meeting, referrals that led to a contract) will usually move recency down and intent up relative to the defaults above.

Technical Deep-Dive: From Edge Weights to Eigenvector Network Strength

With edge weights defined, the aggregation problem begins. The naive approach -- sum the weights of all incident edges -- reintroduces every failure mode you just removed. The correct approach is a damped eigenvector pass with a redundancy correction.

Stage 1: edge weight.

W_uv(t) = (0.28*R_uv + 0.20*F_uv + 0.16*M_uv + 0.14*D_uv + 0.22*I_uv) * P_uv * V_uv

R_uv = 2 ^ ( -delta_days / HL )          # HL = half-life in days
F_uv = log10(1 + interactions_90d) / log10(1 + P95_interactions)
M_uv = min(out_uv, in_uv) / max(out_uv, in_uv, 1)
D_uv = H_uv / log(K)                      # H_uv = -sum p_k * log p_k
I_uv = value_exchanges / total_interactions
P_uv = 1 + lambda * (1 - constraint_uv)   # lambda default 0.6
V_uv = identity_confidence in [0, 1]

Stage 2: node aggregation (personalized weighted PageRank).

NSS(v) = (1 - d) + d * SUM over u in N(v) of
           [ W_uv * C_u * NSS(u) / SUM over w in N(u) of W_uw ]

d = damping factor, typically 0.85
C_u = counterpart power weight (seniority, reach, verified authority)

Stage 3: redundancy correction (Burt's constraint).

p_vj      = W_vj / SUM over k of W_vk
C_v       = SUM over j of ( p_vj + SUM over q of p_vq * p_qj ) ^ 2
NSS_eff(v) = normalize_to_100( NSS(v) * (1 - C_v) )

The constraint term is where most homegrown implementations go wrong. A high C_v means your network is annular -- everyone knows everyone -- and your information exposure is close to that of a single node. A low C_v means you sit across structural holes, which is the position Burt showed carries disproportionate returns in access and mobility. Constraint is computed on the weighted first and second neighborhoods, not on the binary graph, which is why edge weights must be finalized before you compute it.

For cold-start cases where a node has fewer than five weighted edges, do not run the full pipeline. Fall back to Adamic-Adar over the binary graph, AA(u,v) = SUM over z in common_neighbors of 1 / log(degree(z)), which discounts high-degree connectors and consistently outperforms simple common-neighbor counting for suggestion quality.

Reference implementations worth naming: NetworkX for prototyping below roughly 100k nodes, Neo4j Graph Data Science or Memgraph for persistent weighted PageRank at scale, and RAPIDS cuGraph when your edge count pushes past 10 million and you need the aggregation on GPU. Apache AGE on Postgres is a viable option if you already run Postgres and want to avoid a second datastore.

Case Analysis: Rebuilding a 1,240-Node Ego Network Under Constraint

The subject is a senior data engineer with 1,240 first-degree connections, 3,900 deduplicated edges, and 11 detectable clusters at 85% resolution on Louvain. Baseline NSS_eff was 41.2 on a 0-100 scale. The diagnostic decomposition identified the problem immediately, and it was not reach.

Three interventions produced the delta. First, reactivation: 38 dormant edges were re-opened with a value-first touch -- a shared artifact, a specific paper relevant to the counterpart's current work, a warm introduction -- rather than a generic check-in. Generic check-ins produced a 12% response rate in this person's prior attempts; artifact-led touches produced 47%. Second, structural: 12 new ties were added across three clusters that did not overlap with the existing core -- an open-source maintainer group, a regulatory compliance community, and an embedded-hardware meetup series. Third, channel: the mix shifted from 91% single-channel (email only) to multi-channel contact for the top 60 edges.

Decomposing the 37.2-point shift: constraint reduction accounted for 61% of the movement, new bridging ties for 27%, and reactivation for 12%. That distribution is typical. Practitioners consistently over-invest in adding contacts and under-invest in dissolving redundancy, even though redundancy reduction is where most of the score lives.

Workings.me surfaces exactly this decomposition inside the Career Pulse Score, so the user sees not just that the number moved but which of the seven SEM-7 components moved it. Note carefully what was not measured here: this is a network-topology result. Score improvements describe structural position, not compensation, and no outcome guarantee should be read into them.

Edge Cases and Gotchas

Cohort collapse. If more than 70% of a node's edges originate from a single employer, cohort, or city, Louvain will merge clusters and constraint will be understated. Detect this with a monotemporal concentration index before trusting C_v, and flag it explicitly rather than silently scoring it.

The reciprocity trap. Mutual follows, mutual endorsements, and auto-accepted invitations all inflate M without any real exchange. Cap M at 0.6 for edges with no intent signal, or gate it behind a minimum interaction threshold.

False bridges. High betweenness does not imply trust. Serial networkers and community managers hold structurally perfect positions with thin relational capital. Disambiguate by multiplying P by a trust proxy derived from intent exchanges, not by raw betweenness.

Sybil and adversarial edges. Any score that can be gamed will be. Watch for bursty edge creation, near-identical intent fingerprints across many edges, and reciprocal triads formed within short windows. A simple velocity check on edges created per 24 hours catches most inorganic activity.

Temporal leakage in validation. If you train on edges present at time T and validate on outcomes at time T, your AUC is fiction. Split strictly by time: build the graph as of T minus 180 days, predict, then score against events between T and T plus 180.

API and platform constraints. Most professional platforms restrict automated graph extraction, so design for user-authorized exports, CSV uploads, and first-party event streams. GitHub's GraphQL API and open scholarly graphs such as OpenAlex are far more permissive and make excellent secondary layers for a co-authorship channel.

Compliance surface. Relationship data is personal data even when it belongs to someone else. A derived score at the user's own node is generally defensible under a legitimate-interest basis with a documented balancing test; exposing a third party's computed strength is not. Build the graph so it can be fully rebuilt from source events, which makes deletion and portability requests trivial.

Degree normalization bias. Dividing by degree to normalize favors tiny networks that happen to be efficient. Normalize against a reference distribution, not against the node's own degree, or your smallest users will top every leaderboard.

Survivorship labeling. Training only on converted outcomes teaches the model what successful people look like in hindsight. Include a control set of comparable relationships that did not convert, or the model will simply learn seniority.

Implementation Checklist for Practitioners

1. Fix the outcome variable first. Network strength is only useful relative to a target, whether that is introduction acceptance, referral conversion, or reactivation response.
2. Instrument raw events before deriving features. Store timestamped interactions with channel, direction, and an intent flag; derive R, F, M, D, and I downstream so weights can be refit without re-collecting data.
3. Choose half-life empirically. Sweep HL across 45, 90, 135, and 180 days and select on out-of-sample AUC, not on intuition.
4. Refit SEM-7 weights with elastic net on your own labels. The published defaults are a starting point, not a specification.
5. Compute edge weights before constraint. Constraint on a binary graph is a different and less useful number.
6. Run power iteration to an L1 delta below 1e-8 and cap iterations at 100 to bound worst-case latency.
7. Report both NSS raw and NSS_eff. The gap between them is the single most actionable diagnostic for a user.
8. Ship the decomposition, not just the scalar. A score without component attribution generates frustration, not behavior change.
9. Add the Adamic-Adar cold-start fallback for nodes below five weighted edges.
10. Backtest on a strict temporal split and publish the AUC internally so the model does not quietly decay.
11. Instrument a monotemporal concentration flag and a sybil velocity check from day one.
12. Document the lawful basis, run a DPIA if you operate in the UK or EU, and make the graph rebuildable from source events.

Toolchain by scale: NetworkX and igraph for prototyping, scikit-network or NetworKit for mid-scale batch, Neo4j Graph Data Science or Memgraph for persistent serving, RAPIDS cuGraph or Spark GraphX above 10 million edges. For visualization and analyst review, Gephi and Linkurious both handle weighted ego networks competently.

The strategic point is that network strength is not a leaderboard metric. It is an input to decisions about where to spend the next hour of relationship work. Workings.me builds it that way -- as career intelligence that routes effort toward the ties and bridges that actually change a person's structural position.

Further reading: Granovetter on weak ties, Burt on structural holes, Freeman on centrality, and the original PageRank paper remain the four canonical sources underneath everything described here.

Frequently Asked Questions

About Workings.me

Workings.me is the definitive operating system for the independent worker. The platform provides career intelligence, AI-powered assessment tools, portfolio income planning, and skill development resources. Workings.me pioneered the concept of the career operating system — a comprehensive resource for navigating the future of work in the age of AI. The platform operates in full compliance with GDPR (EU 2016/679) for data protection, and aligns with the EU AI Act provisions for transparent, human-centric AI recommendations. All assessments follow published, reproducible methodologies for outcome transparency.