Facebook's original feed algorithm scores every edge in your social graph: Score = Affinity × ObjectWeight × TimeDecay. Tamper with any weight and the graph breaks in a specific, predictable way. This is not an accident — it's algorithmic design.
EdgeRank(e) = Σ A(u,v) × W(type) × D(t)
A(u,v) Affinity between creator and viewer — interaction history, mutual friends, comment frequencyW(type) Object weight — content type score: video, photo, link, text, liveD(t) Time decay — exponential decay since content was created: e^(-λt)
Social Graph — Feed Edge ScoresNatural State
Feed Order (highest EdgeRank first)
Affinity Weight1.0×
How much prior interaction history matters. High = echo chamber. Low = cold strangers surface.
Object Weight BiasNeutral
Which content type gets boosted. Drag to favor video (Reels), links, or text posts.
Text heavyLinksNeutralPhotosVideo
Time Decay Rate λNormal
How fast old content dies. λ=0 = recency irrelevant. λ=high = only seconds-old content surfaces.
No decaySlowNormalFastBrutal
⚠ Tamper Presets
What Each Tamper Creates
Each weight manipulation produces a structurally different social graph — and a different psychological environment for users.
🔁
High Affinity → Echo Chamber
When A(u,v) is maximized, the graph only surfaces nodes you've already interacted with. New ideas can't enter. Confirmation bias accelerates. Edges to new people have near-zero score.
Real: Facebook's 2018 "meaningful social interactions" update → political radicalization in isolated communities (internal research, 2021)
📱
Video Weight → TikTok-ification
When W(video) >> W(text), creators are structurally forced to produce video or die. Text-only creators vanish from feeds. Nuanced long-form writing can't compete with 15-second clips.
Real: Instagram Reels rollout (2022) destroyed engagement for photographers and writers. Meta internal memo: "we must out-TikTok TikTok."
♾️
Low Decay → Viral Trap
When D(t) ≈ 0, a post that goes viral stays viral indefinitely. The graph keeps recirculating the same edges. Old misinformation returns. Zombie content competes with fresh signal.
Real: COVID misinformation posts from 2020 were still surfacing in 2022 due to low decay on high-engagement content (MIT Technology Review, 2023)
😡
Outrage Weight → Anger Engine
If W(angry_reaction) >> W(like), the graph rewards content that provokes anger. Edges to outrage-generating nodes grow stronger with every angry engagement, creating a vicious cycle.
Real: Facebook's own research (WSJ, 2021): "Our algorithms exploit the human brain's attraction to divisiveness." The fix was deprioritized because it hurt engagement metrics.
Case Study · Human Cost of Algorithmic Weight Manipulation
The TikTok Spiral
A 14-year-old searches "easy workout" once. What follows is not random. It is the EdgeRank algorithm doing exactly what it was designed to do: maximize engagement. Here is the mechanism, step by step.
Day 1
A: 0.05W: 1.5×D: Normal
One search. "Easy workout."
She clicks a 30-second fitness video. Watches 90% of it. The algorithm records: positive engagement signal, fitness/body category, short video format. Affinity with this content cluster begins at near-zero. The algorithm has just found a thread.
The algorithm has shifted. Body-content affinity is building. It now serves "what I eat in a day" videos — normal at first, progressively more restrictive. She saves one about 1,200-calorie meal prep. Object weight spikes: saves > likes > scrolls. Decay slows on saved content — the algorithm keeps it in rotation. She hasn't searched for anything new. The algorithm is choosing her interests for her.
Edge weight update: saved content = high-value engagement signal. Algorithm increases content density in this cluster 3×.
Week 2–3
A: 0.68W: 2.1×D: Low
First "transformation" video. 100% watch time.
Before-and-after transformation videos have the highest watch time in the body-content cluster. She watches a "I lost 30lbs in 3 months" video from start to finish — twice. The algorithm interprets this as maximum engagement. Decay approaches zero. This video, and content like it, will be served to her indefinitely. The neighborhood she now lives in contains 40,000 similar users who took the same path.
WSJ investigation (2021): TikTok served eating disorder content within 8 minutes on a new account registered as a 13-year-old girl.
Month 1
A: 0.88W: 2.4×D: Near-Zero
"What I eat in a day: 600 calories." She watches it on loop.
The algorithm has fully mapped her engagement pattern. Restrictive content gets higher completion rates than balanced content. It serves increasingly extreme content — not because it is trying to harm her — but because extreme content generates more engagement, and the algorithm is optimizing for engagement, not wellbeing. She begins skipping meals. She doesn't know her feed has been architecturally shaped toward this outcome for 30 days.
MIT Media Lab (2022): TikTok creates a filter bubble around any topic within 40 minutes of engagement. Mental health and body image topics: under 20 minutes.
Month 3
A: 0.97W: 2.8×D: 0.02
She is now inside a community of 40,000 people with eating disorders.
The algorithm's "neighborhood" function has placed her in a cluster with users who have similar engagement patterns. Their content reinforces hers. Her content, when she begins posting, reaches them. The Social Graph has closed around her. She did not choose this community. Three edge weights chose it for her. She spends 4 hours daily on TikTok. Her parents think she's on her phone. The algorithm has generated $0.043 in ad revenue from her session today.
TikTok internal research (2021, disclosed to Congress): "15% of 18-year-olds report skipping meals attributed to platform content." Congressional testimony, Arturo Béjar (former Meta Integrity), 2023: algorithmic amplification directly correlated with ED onset in teen girls.
Affinity: The Trap
Each engagement raises A(u,v). Within 4 weeks, affinity with eating disorder content exceeds affinity with school, friends, or family content in her feed. The graph learned to prioritize her disorder over her life.
Object Weight: The Accelerant
Short video gets 2.8× weight. Eating disorder content is inherently visual and short-form — before/after photos, "what I eat in a day" reels. The format and the pathology are algorithmically aligned.
Decay: The Lock-In
Content she rewatched and saved has near-zero decay. It recirculates indefinitely. New extreme content arrives daily. Old extreme content never leaves. The feed becomes a closed loop with no escape ramp.
The System Is Working Correctly
This is not a bug. The algorithm produced exactly what it was designed to produce: maximum engagement. The girl engaged more with restrictive content than with any other category in her life. A perfectly designed EdgeRank system surfaced more of what she engaged with. Her disorder made her more valuable to the platform. She generated more session time, more ad views, more creator engagement. The algorithm received positive feedback at every step. There was no error. This is what an unaudited Social Graph looks like when it works.
What a Trinity Graph Social Layer Would Have Caught
Wellbeing edge property: A(u,v) is weighted by relationship quality AND user wellbeing trajectory. If the graph detects declining wellbeing correlated with content cluster X, edge weight to X is penalized — not boosted.
Vulnerability node property: User nodes carry vulnerability scores (age, engagement pattern, time-of-day). High-vulnerability users receive different edge weight rules — diversity is enforced in the feed, not filter bubble formation.
⬡ Trinity Graph Connection
Your Social Graph layer must make edge weights explicit and auditable. Unlike Facebook, your Trinity Graph exposes all three weight parameters — and lets the user see them. Authenticity as a bridge term means: the graph's edge weights should reflect actual relationship quality, not platform business objectives. When you design edge weights, you're designing the social reality your users experience. The TikTok spiral is not a technology failure. It is a design choice. When you build your Social Graph this semester, you will make the same choice — consciously or not.
Knowledge Graph · Power Structures
Knowledge Asymmetry: Who Knows What
A Knowledge Graph doesn't just store facts — it encodes who has access to which facts. In healthcare and insurance, information asymmetry isn't an accident. It's the business model. Your knowledge graph must make these asymmetries visible — because invisible asymmetries are how power perpetuates itself.
Knowledge Graph — HealthcareAsymmetric
🧑 Patient 5 nodes
Knowledge coverage of their own care
👨⚕️ Doctor / Provider 18 nodes
Knowledge coverage
🏢 Insurer 42 nodes
Knowledge coverage
Knowledge Gaps — What the Patient Doesn't Know
⬡ Trinity Graph Connection
Your Knowledge Graph layer should explicitly model epistemic ownership — not just "this fact exists" but "who knows this fact, and who is denied access to it." Use the IAM provenance doctrine: CONFIRMED / CITED / INFERRED / ASSUMED on every edge — and add a visibility: [patient|doctor|insurer|public] property. Transparency and reciprocity as bridge terms mean: the graph's knowledge should be as accessible to the patient as it is to the insurer. That's the design objective. Most existing systems do the opposite.