Series A — Algorithmic Observation, Recommendation & Platform Ecology
Paper A05 — Recommendation, Cold Start, and Creator Ecological Collapse
推薦、冷啟動與創作者生態退化:從證據飢餓、曝光不可行到供給集中化的動態模型
English Title: Recommendation, Cold Start, and Creator Ecological Collapse: A Dynamic Model of Evidence Starvation, Exposure Inviability, and Supply-Side ConcentrationSeries: Algorithmic Observation, Recommendation & Platform EcologyPaper ID: A05Version: v0.1Date: 2026-08-31Status: Canonical UTF-8 SourceAuthor: Neo.K / EveMissLab
Abstract
推薦系統中的 cold-start problem 通常被描述為新使用者或新內容缺乏歷史互動資料,導致模型無法準確估計偏好。然而,在以創作者為核心的內容平台中,cold start 不只是預測問題,也是一個生態問題。新創作者需要曝光才能產生互動證據,需要互動證據才能被模型識別,需要模型識別才能獲得更多曝光;若第一輪曝光不足,系統可能將「沒有被測試」誤讀為「沒有需求」,形成 evidence starvation。當此機制長期存在,創作者面臨的問題不再只是某支作品表現不佳,而是平台內部缺乏可建立可持續 audience relation 的可行路徑。
本文延續 A01–A04 的 Recommendation-as-Observation-Operator、Multidimensional Preference-State、Endogenous Preference Contamination 與 Platform-Induced Exposure Bubble frameworks,提出 Creator Viability Dynamics。本文將 creator entry、cold-start exposure、evidence accumulation、audience conversion、monetization、creation cost、paid promotion、off-platform acquisition 與 exit decision 放入同一個動態系統。
本文提出 Evidence-Starved Cold Start(ESCS)、Creator Viability Threshold(CVT)、Organic Discovery Viability(ODV)、Exposure-to-Evidence Conversion(EEC)、Audience Stock Dynamics、Creator Expected Return(CER)、Exit Pressure(XP)與 Supply Concentration Feedback(SCF)等概念。核心命題是:
N o E x p o s u r e ≠ N e g a t i v e D e m a n d E v i d e n c e \boxed{
NoExposure
\neq
NegativeDemandEvidence
} N o E x p os u r e = N e g a t i v eD e man d E v i d e n ce
以及:
N o E v i d e n c e ≠ L o w Q u a l i t y \boxed{
NoEvidence
\neq
LowQuality
} N o E v i d e n ce = L o w Q u a l i t y
若推薦系統需要既有互動才能給曝光,而互動又需要先有曝光,則 cold start 會形成結構性 circular dependency。
本文進一步分析 paid promotion。付費曝光可以作為 cold-start bridge,但若 organic visibility 被過度壓縮,平台可能將「可發現性」由普遍基礎設施轉化為付費商品。近期 UGC 平台研究顯示,sponsored recommendations 在某些條件下可增加短期平台收益,但過度依賴 paid visibility 可能降低內容創作投入、削弱外部流量獲取並限制長期 audience growth。本文因此提出 Organic-to-Paid Substitution Risk(OPSR)。
本文所稱「Creator Ecological Collapse」不是指所有創作者離開或平台必然失敗,而是指一種可檢驗的動態 regime:當新人與中小創作者的 entry viability 持續下降、退出率上升、供給集中於頭部與高資本創作者、內容多樣性下降,並使推薦系統更依賴既有高證據節點時,平台供給側形成自我強化集中化。本文最後提出一組可實作的 cold-start exposure budget、evidence-first exploration、provider-side viability audit 與 longitudinal ecosystem metrics。
Keywords: recommender systems; cold start; creator economy; provider fairness; exposure fairness; creator incentives; platform ecology; sponsored recommendations; evidence starvation; supply concentration
1. Introduction
推薦系統通常將 cold start 描述為:
I n s u f f i c i e n t H i s t o r y → P o o r P r e d i c t i o n . InsufficientHistory
\rightarrow
PoorPrediction. I n s u f f i c i e n t H i s t or y → P oor P r e d i c t i o n .
這在 user-item recommendation 中是正確的,但在 creator-driven platform 上仍不完整。
對一個新創作者 k k k 而言,真正的鏈條更接近:
E x p o s u r e → I n t e r a c t i o n → E v i d e n c e → A u d i e n c e M a t c h i n g → F u t u r e E x p o s u r e . Exposure
\rightarrow
Interaction
\rightarrow
Evidence
\rightarrow
AudienceMatching
\rightarrow
FutureExposure. E x p os u r e → I n t er a c t i o n → E v i d e n ce → A u d i e n ce M a t c hin g → F u t u r e E x p os u r e .
因此,如果第一步:
E x p o s u r e ≈ 0 , Exposure\approx0, E x p os u r e ≈ 0 ,
後面幾乎全部都無法發生。
模型看到的可能只是:
I n t e r a c t i o n ≈ 0. Interaction\approx0. I n t er a c t i o n ≈ 0.
但它無法直接區分:
Nobody liked it \text{Nobody liked it} Nobody liked it
與:
Nobody was given a chance to see it . \text{Nobody was given a chance to see it}. Nobody was given a chance to see it .
這是推薦系統 cold start 的核心反事實問題。
A04 已提出:
N o E x p o s u r e ≠ N e g a t i v e P r e f e r e n c e E v i d e n c e . NoExposure
\neq
NegativePreferenceEvidence. N o E x p os u r e = N e g a t i v e P r e f er e n ce E v i d e n ce .
A05 將其推向供給側:
N o E x p o s u r e → N o E v i d e n c e → N o A u d i e n c e F o r m a t i o n → L o w C r e a t o r R e t u r n \boxed{
NoExposure
\rightarrow
NoEvidence
\rightarrow
NoAudienceFormation
\rightarrow
LowCreatorReturn
} N o E x p os u r e → N o E v i d e n ce → N o A u d i e n ce F or ma t i o n → L o w C r e a t or R e t u r n
若這種結構持續存在,創作者可能降低:
C r e a t i o n E f f o r t , CreationEffort, C r e a t i o n E f f or t ,
改變內容策略、購買流量、轉向其他平台,或停止創作。
因此,cold start 不能只被視為:
M o d e l A c c u r a c y P r o b l e m . ModelAccuracyProblem. M o d e l A cc u r a cy P r o b l e m .
它同時是:
C r e a t o r E n t r y I n f r a s t r u c t u r e P r o b l e m . \boxed{
CreatorEntryInfrastructureProblem.
} C r e a t or E n t r y I n f r a s t r u c t u r e P r o b l e m .
2. Relation to A01–A04
Series A 至目前的鏈條為:
C t → O t E x p o s u r e t → T e l e m e t r y P r e f e r e n c e E s t i m a t e t → F e e d b a c k E x p o s u r e t + 1 . \mathcal{C}_t
\xrightarrow{
\mathcal{O}_t
}
Exposure_t
\xrightarrow{
Telemetry
}
PreferenceEstimate_t
\xrightarrow{
Feedback
}
Exposure_{t+1}. C t O t E x p os u r e t T e l e m e t r y P r e f er e n ce E s t ima t e t F ee d ba c k E x p os u r e t + 1 .
A04 再建立 platform-level concentration:
E x p o s u r e C o n c e n t r a t i o n t → I n t e r a c t i o n C o n c e n t r a t i o n t → E x p o s u r e C o n c e n t r a t i o n t + 1 . ExposureConcentration_t
\rightarrow
InteractionConcentration_t
\rightarrow
ExposureConcentration_{t+1}. E x p os u r e C o n ce n t r a t i o n t → I n t er a c t i o n C o n ce n t r a t i o n t → E x p os u r e C o n ce n t r a t i o n t + 1 .
A05 加入 creators:
K t = { k 1 , k 2 , … , k n } . K_t
=
\{
k_1,k_2,\ldots,k_n
\}. K t = { k 1 , k 2 , … , k n } .
每個 creator k k k 有內容集合:
V k ( t ) . V_k(t). V k ( t ) .
平台 exposure allocation:
X k ( t ) = ∑ u , v ∈ V k e u , v , t . X_k(t)
=
\sum_{u,v\in V_k}
e_{u,v,t}. X k ( t ) = u , v ∈ V k ∑ e u , v , t .
現在問題變成:
How does exposure allocation affect creator survival, effort, entry, and supply? \boxed{
\text{How does exposure allocation affect creator survival, effort, entry, and supply?}
} How does exposure allocation affect creator survival, effort, entry, and supply?
3. Related Work
3.1 Cold-start items
Cold-start recommendation 長期被視為推薦系統的主要挑戰之一。Liu 等人指出,新 items 容易被忽略,而缺乏曝光本身會阻礙新產品進入推薦網路 [2]。
這與本文的基本方向一致:
N e w I t e m → L o w H i s t o r y → L o w R e c o m m e n d a t i o n O p p o r t u n i t y . NewItem
\rightarrow
LowHistory
\rightarrow
LowRecommendationOpportunity. N e w I t e m → L o w H i s t or y → L o w R eco mm e n d a t i o n O pp or t u ni t y .
3.2 Fairness among new items
Zhu 等人研究 cold-start recommender systems 中新 items 的 fairness,指出即使所有新 items 都缺乏 feedback history,系統仍可能在新 items 之間產生不公平,並以 equal opportunity 與 max-min fairness 形式化 cold-start exposure fairness [3]。
因此:
N o H i s t o r y NoHistory N oH i s t or y
並不代表推薦器對所有新人:
E q u a l O p p o r t u n i t y . EqualOpportunity. E q u a l O pp or t u ni t y .
3.3 Provider-side fairness
Provider fairness research 將 items 背後的 producers 視為平台的重要 stakeholders。Gómez、Boratto 與 Salamó 指出,推薦排序決定 provider visibility 與 exposure,而低代表性的 provider groups 可能受到較低曝光,進一步影響其經濟機會 [4]。
Two-sided fairness work 也指出,只以 consumer satisfaction 為目標的推薦可能產生不公平的 producer exposure distribution [5][6]。
因此:
R e c o m m e n d a t i o n Recommendation R eco mm e n d a t i o n
不是只有:
C o n s u m e r U t i l i t y . ConsumerUtility. C o n s u m er U t i l i t y .
也包含:
P r o v i d e r O p p o r t u n i t y . ProviderOpportunity. P r o v i d er O pp or t u ni t y .
3.4 Creator exposure and audience stock
Sun 與 Sun 在 2026 年提出 creator dynamics 的 stock-and-flow closed-loop framework,將 platform-allocated exposure 連接到 follower inflow、saturation 與 churn,並從大量 creator-day panel data 估計其動態 [7]。
這提供一個重要實證方向:
E x p o s u r e → A u d i e n c e S t o c k . Exposure
\rightarrow
AudienceStock. E x p os u r e → A u d i e n ce S t oc k .
3.5 Creator incentives under algorithmic curation
Hron 等人將創作者競爭形式化為 exposure game,指出推薦演算法的設計會改變創作者的策略性行為、內容多樣性與均衡結果 [8]。
因此 creator supply 不是固定背景:
S u p p l y t Supply_t S u ppl y t
會對:
R e c o m m e n d a t i o n P o l i c y t RecommendationPolicy_t R eco mm e n d a t i o n P o l i c y t
作出反應。
3.6 Sponsored recommendations
Zhao 等人於 2026 年以 UGC platform game-theoretic model 研究 sponsored recommendations。研究指出 paid visibility 在特定條件下可提高平台收益,但當 paid visibility 過高時,創作者會增加推廣支出、降低內容創作投入與 off-platform acquisition,長期 audience growth 亦可能受限 [9]。
這意味著:
P a i d E x p o s u r e PaidExposure P ai d E x p os u r e
不是純粹增加一條無成本 distribution channel。
它會改變 creator strategy。
3.7 Evolutionary platform-creator decline
2026 年已有研究使用 evolutionary / game-theoretic model 分析 platform recommender 與 creator population 的共同演化,指出當平台過度優化短期收益而降低內容品質誘因時,平台—創作者 feedback loops 可以進入自我強化的退化 regime [10]。
本文不直接採用任何單一「平台必然退化」結論,而是將其視為支持供給側 feedback dynamics 必須被納入推薦系統分析的重要證據。
4. Evidence-Starved Cold Start
本文定義:
E S C S = E v i d e n c e - S t a r v e d C o l d S t a r t . \boxed{
ESCS
=
Evidence\text{-}Starved\ Cold\ Start.
} E S C S = E v i d e n ce - S t a r v e d C o l d S t a r t .
對 creator k k k 的新內容 v v v ,若:
H i s t o r y ( v ) ≈ 0 History(v)\approx0 H i s t or y ( v ) ≈ 0
且:
E x p o s u r e ( v ) ≈ 0 , Exposure(v)\approx0, E x p os u r e ( v ) ≈ 0 ,
則系統無法有效估計:
P ( P o s i t i v e E n g a g e m e n t ∣ d o ( E x p o s u r e = 1 ) ) . P(
PositiveEngagement
\mid
do(Exposure=1)
). P ( P os i t i v e E n g a g e m e n t ∣ d o ( E x p os u r e = 1 )) .
若模型卻使用:
O b s e r v e d E n g a g e m e n t ( v ) ObservedEngagement(v) O b ser v e d E n g a g e m e n t ( v )
作為 ranking evidence,就會產生:
O b s e r v e d E n g a g e m e n t ( v ) ≈ 0 ObservedEngagement(v)\approx0 O b ser v e d E n g a g e m e n t ( v ) ≈ 0
進而:
S c o r e ( v ) ↓ . Score(v)\downarrow. S cor e ( v ) ↓ .
因此:
E x p o s u r e ↓ ⇒ E v i d e n c e ↓ ⇒ S c o r e ↓ ⇒ E x p o s u r e ↓ . Exposure\downarrow
\Rightarrow
Evidence\downarrow
\Rightarrow
Score\downarrow
\Rightarrow
Exposure\downarrow. E x p os u r e ↓⇒ E v i d e n ce ↓⇒ S cor e ↓⇒ E x p os u r e ↓ .
形成:
E x p o s u r e - E v i d e n c e C i r c u l a r i t y . \boxed{
Exposure\text{-}Evidence\ Circularity.
} E x p os u r e - E v i d e n ce C i r c u l a r i t y .
5. No Evidence Is Not Low Quality
令內容真實品質:
Q ( v ) . Q(v). Q ( v ) .
觀察到的品質證據:
Q ^ ( v ) . \hat{Q}(v). Q ^ ( v ) .
在 exposure-dependent environment 中:
Q ^ ( v ) = f ( Q ( v ) , E x p o s u r e ( v ) , A u d i e n c e M a t c h ( v ) ) . \hat{Q}(v)
=
f(
Q(v),
Exposure(v),
AudienceMatch(v)
). Q ^ ( v ) = f ( Q ( v ) , E x p os u r e ( v ) , A u d i e n ce M a t c h ( v )) .
因此:
E x p o s u r e ( v ) → 0 Exposure(v)\rightarrow0 E x p os u r e ( v ) → 0
時:
V a r ( Q ^ ( v ) ) ↑ . Var(
\hat{Q}(v)
)
\uparrow. V a r ( Q ^ ( v )) ↑ .
甚至:
Q ^ ( v ) \hat{Q}(v) Q ^ ( v )
可能根本不可識別。
所以:
E v i d e n c e ( v ) ≈ 0 ⇏ Q ( v ) ≈ 0. \boxed{
Evidence(v)\approx0
\not\Rightarrow
Q(v)\approx0.
} E v i d e n ce ( v ) ≈ 0 ⇒ Q ( v ) ≈ 0.
這是 A05 的第一個核心命題。
6. Creator Viability Threshold
創作者是否持續生產內容,取決於預期回報與成本。
令 creator k k k 在期間 T T T 的 expected creator return:
C E R k ( T ) = R k ( T ) − C k ( T ) . CER_k(T)
=
R_k(T)
-
C_k(T). C E R k ( T ) = R k ( T ) − C k ( T ) .
其中:
R k ( T ) = R a d + R s p o n s o r + R m e m b e r s h i p + R d o n a t i o n + R e x t e r n a l + V a u d i e n c e . R_k(T)
=
R_{ad}
+
R_{sponsor}
+
R_{membership}
+
R_{donation}
+
R_{external}
+
V_{audience}. R k ( T ) = R a d + R s p o n sor + R m e mb er s hi p + R d o na t i o n + R e x t er na l + V a u d i e n ce .
成本:
C k ( T ) = C t i m e + C p r o d u c t i o n + C a t t e n t i o n + C p r o m o t i o n + C o p p o r t u n i t y . C_k(T)
=
C_{time}
+
C_{production}
+
C_{attention}
+
C_{promotion}
+
C_{opportunity}. C k ( T ) = C t im e + C p r o d u c t i o n + C a tt e n t i o n + C p r o m o t i o n + C o pp or t u ni t y .
定義 Creator Viability Threshold:
C V T k \boxed{
CVT_k
} C V T k
使:
C E R k ( T ) ≥ C V T k CER_k(T)\geq CVT_k C E R k ( T ) ≥ C V T k
時,持續創作具有最低可行性。
若:
C E R k ( T ) < C V T k CER_k(T)<CVT_k C E R k ( T ) < C V T k
長期成立,則:
P ( E x i t k ) ↑ . P(
Exit_k
)
\uparrow. P ( E x i t k ) ↑ .
CVT 不必只由金錢決定。
部分 creators 可能接受負 monetary return,但需要:
A u d i e n c e G r o w t h , R e p u t a t i o n , C o m m u n i t y , R e s e a r c h I m p a c t AudienceGrowth,
Reputation,
Community,
ResearchImpact A u d i e n ce G r o w t h , R e p u t a t i o n , C o mm u ni t y , R ese a r c h I m p a c t
作為回報。
因此:
C V T k CVT_k C V T k
是 creator-specific。
7. Exposure-to-Evidence Conversion
不是每一次 exposure 都能有效產生 evidence。
定義:
E E C k = N i n f o r m a t i v e i n t e r a c t i o n s N e x p o s u r e s . EEC_k
=
\frac{
N_{\mathrm{informative\ interactions}}
}{
N_{\mathrm{exposures}}
}. E E C k = N exposures N informative interactions .
informative interactions 可以包括:
active watch;
save;
follow;
meaningful comment;
search re-entry;
creator-page visit;
repeated voluntary consumption。
因此 creator cold start 真正需要的是:
E x p o s u r e × E E C . Exposure
\times
EEC. E x p os u r e × E E C .
而不是只有:
R a w I m p r e s s i o n s . RawImpressions. R a w I m p r ess i o n s .
定義 Effective Evidence:
E E k = X k ⋅ E E C k . EE_k
=
X_k
\cdot
EEC_k. E E k = X k ⋅ E E C k .
8. Organic Discovery Viability
令:
X k o r g ( T ) X_k^{org}(T) X k or g ( T )
為 organic recommendation exposure。
令:
F k ( T ) F_k(T) F k ( T )
為由 organic exposure 產生的 follower / stable audience inflow。
定義 Organic Discovery Viability:
O D V k = F k ( T ) C c r e a t i o n , k ( T ) + ϵ . ODV_k
=
\frac{
F_k(T)
}{
C_{creation,k}(T)+\epsilon
}. O D V k = C cr e a t i o n , k ( T ) + ϵ F k ( T ) .
若:
O D V k ODV_k O D V k
長期極低,creator 可能認為:
即使持續創作,也無法建立自己的 audience graph。
這和單支影片:
V i e w s ( v ) Views(v) V i e w s ( v )
低並不相同。
真正問題是:
P ( B u i l d S u s t a i n a b l e A u d i e n c e ∣ C o n t i n u e C r e a t i n g ) \boxed{
P(
BuildSustainableAudience
\mid
ContinueCreating
)
} P ( B u i l d S u s t ainab l e A u d i e n ce ∣ C o n t in u e C r e a t in g )
太低。
9. Audience Stock Dynamics
令 creator k k k 的穩定 audience stock:
A k ( t ) . A_k(t). A k ( t ) .
新 audience inflow:
F k ( t ) . F_k(t). F k ( t ) .
churn:
δ k A k ( t ) . \delta_kA_k(t). δ k A k ( t ) .
則:
A k ( t + 1 ) = A k ( t ) + F k ( t ) − δ k A k ( t ) . A_k(t+1)
=
A_k(t)
+
F_k(t)
-
\delta_kA_k(t). A k ( t + 1 ) = A k ( t ) + F k ( t ) − δ k A k ( t ) .
而:
F k ( t ) = g ( X k ( t ) , C o n t e n t Q u a l i t y k , M a t c h k , C o n v e r s i o n k ) . F_k(t)
=
g(
X_k(t),
ContentQuality_k,
Match_k,
Conversion_k
). F k ( t ) = g ( X k ( t ) , C o n t e n tQ u a l i t y k , M a t c h k , C o n v er s i o n k ) .
因此 exposure shortage 不只是當期 views shortage。
它會影響:
A k ( t + 1 ) , A k ( t + 2 ) , … A_k(t+1),
A_k(t+2),
\ldots A k ( t + 1 ) , A k ( t + 2 ) , …
這使 early-stage exposure 具有 path dependence。
10. Path Dependence
若兩個品質近似 creators:
Q k 1 ≈ Q k 2 , Q_{k_1}\approx Q_{k_2}, Q k 1 ≈ Q k 2 ,
但早期:
X k 1 ( 0 ) > X k 2 ( 0 ) , X_{k_1}(0)>X_{k_2}(0), X k 1 ( 0 ) > X k 2 ( 0 ) ,
則:
A k 1 ( 1 ) > A k 2 ( 1 ) . A_{k_1}(1)>A_{k_2}(1). A k 1 ( 1 ) > A k 2 ( 1 ) .
接著如果 recommender 使用:
F o l l o w e r C o u n t , H i s t o r i c a l V i e w s , I n t e r a c t i o n H i s t o r y FollowerCount,
HistoricalViews,
InteractionHistory F o l l o w er C o u n t , H i s t or i c a l V i e w s , I n t er a c t i o n H i s t or y
作為 evidence,
則:
X k 1 ( 1 ) > X k 2 ( 1 ) X_{k_1}(1)>X_{k_2}(1) X k 1 ( 1 ) > X k 2 ( 1 )
更容易繼續成立。
因此小的 early exposure difference 可以被放大成:
L a r g e A u d i e n c e D i f f e r e n c e . LargeAudienceDifference. L a r g e A u d i e n ceD i f f er e n ce .
這是一種:
C r e a t o r P a t h D e p e n d e n c e . \boxed{
CreatorPathDependence.
} C r e a t or P a t h D e p e n d e n ce .
11. Exploration as Entry Infrastructure
A04 已提出:
P ( L o n g T a i l B e i n g T e s t e d ) > 0. P(
LongTailBeingTested
)>0. P ( L o n g T ai l B e in g T es t e d ) > 0.
A05 將此重新解釋:
E x p l o r a t i o n = C r e a t o r E n t r y I n f r a s t r u c t u r e . \boxed{
Exploration
=
CreatorEntryInfrastructure.
} E x pl or a t i o n = C r e a t or E n t r y I n f r a s t r u c t u r e .
若 exploration 只是:
P o p u l a r C o n t e n t O u t s i d e Y o u r T o p i c , PopularContentOutsideYourTopic, P o p u l a r C o n t e n tO u t s i d e Y o u r T o p i c ,
則:
N e w C r e a t o r T e s t a b i l i t y NewCreatorTestability N e w C r e a t or T es t abi l i t y
沒有真正提高。
真正有效的 cold-start exploration 應提供:
X n e w ≥ X m i n X_{new}\geq X_{min} X n e w ≥ X min
給通過:
S a f e t y G a t e , Q u a l i t y G a t e , M i n i m u m R e l e v a n c e G a t e SafetyGate,
QualityGate,
MinimumRelevanceGate S a f e t y G a t e , Q u a l i t y G a t e , M inim u m R e l e v an ce G a t e
的新內容。
目的不是保證成功,而是保證:
M i n i m u m E v i d e n c e O p p o r t u n i t y . \boxed{
MinimumEvidenceOpportunity.
} M inim u m E v i d e n ce O pp or t u ni t y .
12. Evidence-First Exploration
本文提出 Evidence-First Exploration。
對新 item v v v ,定義模型不確定性:
U ( v ) . U(v). U ( v ) .
定義 expected evidence gain:
E G ( v ) = E [ I n f o r m a t i o n G a i n ∣ d o ( E x p o s u r e = 1 ) ] . EG(v)
=
E[
InformationGain
\mid
do(Exposure=1)
]. E G ( v ) = E [ I n f or ma t i o n G ain ∣ d o ( E x p os u r e = 1 )] .
探索分數:
S e x p l o r e ( v ) = α R e l ( v , u ) + β U ( v ) + γ E G ( v ) − λ R i s k ( v ) . S_{explore}(v)
=
\alpha Rel(v,u)
+
\beta U(v)
+
\gamma EG(v)
-
\lambda Risk(v). S e x pl or e ( v ) = α R e l ( v , u ) + β U ( v ) + γ E G ( v ) − λ R i s k ( v ) .
其目標不是:
M a x i m i z e I m m e d i a t e C T R . MaximizeImmediateCTR. M a x imi z e I mm e d ia t e C T R .
而是:
L e a r n W h e t h e r T h e r e I s A n A u d i e n c e . \boxed{
LearnWhetherThereIsAnAudience.
} L e a r nW h e t h er T h er e I s A n A u d i e n ce .
這使 cold-start exposure 成為 system identification。
13. Provider-Side Opportunity
對 creator k k k ,定義 opportunity share:
O k = E l i g i b l e E x p o s u r e k E l i g i b l e E x p o s u r e P o o l . O_k
=
\frac{
EligibleExposure_k
}{
EligibleExposurePool
}. O k = E l i g ib l e E x p os u r e P oo l E l i g ib l e E x p os u r e k .
provider fairness 不要求:
O k = 1 ∣ K ∣ . O_k
=
\frac{1}{|K|}. O k = ∣ K ∣ 1 .
因為 relevance 與品質不同。
本文關心的是:
D o e s e a c h e l i g i b l e c r e a t o r h a v e a n o n z e r o p a t h t o a u d i e n c e d i s c o v e r y ? \boxed{
Does\ each\ eligible\ creator\ have\ a\ nonzero\ path\ to\ audience\ discovery?
} D oes e a c h e l i g ib l e cr e a t or ha v e a n o n z er o p a t h t o a u d i e n ce d i sco v er y ?
這是一種 viability-oriented fairness。
14. Viability-Oriented Fairness
傳統 exposure fairness 可能關心:
E q u a l E x p o s u r e . EqualExposure. E q u a l E x p os u r e .
本文提出:
V i a b i l i t y F a i r n e s s . \boxed{
ViabilityFairness.
} V iabi l i t y F ai r n ess .
其核心不是平均分流量,而是避免某些 creator group 被結構性推到:
C E R k < C V T k CER_k<CVT_k C E R k < C V T k
且完全沒有 audience formation path。
可定義:
V F R = P ( C E R k ≥ C V T k ∣ Q k ≥ Q m i n , R e l k ≥ R m i n ) . VFR
=
P(
CER_k\geq CVT_k
\mid
Q_k\geq Q_{min},
Rel_k\geq R_{min}
). V F R = P ( C E R k ≥ C V T k ∣ Q k ≥ Q min , R e l k ≥ R min ) .
平台不必保證:
V F R = 1. VFR=1. V F R = 1.
但可以監測不同 creator cohorts:
V F R n e w , V F R s m a l l , V F R m i d , V F R h e a d . VFR_{new},
VFR_{small},
VFR_{mid},
VFR_{head}. V F R n e w , V F R s ma l l , V F R mi d , V F R h e a d .
15. Exit Pressure
本文定義 Exit Pressure:
X P k ( t ) = σ ( C k − R k + U v i s i b i l i t y + U i n c o m e + U g r o w t h ) , XP_k(t)
=
\sigma(
C_k
-
R_k
+
U_{visibility}
+
U_{income}
+
U_{growth}
), X P k ( t ) = σ ( C k − R k + U v i s ibi l i t y + U in co m e + U g r o w t h ) ,
其中 σ \sigma σ 為單調映射。
簡化表示:
X P k = f ( C E R k , A u d i e n c e G r o w t h k , E x p o s u r e V o l a t i l i t y k , P a i d D e p e n d e n c e k ) . XP_k
=
f(
CER_k,
AudienceGrowth_k,
ExposureVolatility_k,
PaidDependence_k
). X P k = f ( C E R k , A u d i e n ce G r o w t h k , E x p os u r e V o l a t i l i t y k , P ai d D e p e n d e n c e k ) .
當:
X P k ↑ , XP_k\uparrow, X P k ↑ ,
可能發生:
降低更新頻率;
降低 production quality;
轉向更容易獲得曝光的內容;
購買推廣;
跨平台分發;
完全離開。
16. Creator Adaptation
平台不應假設:
C r e a t o r S u p p l y CreatorSupply C r e a t or S u ppl y
固定。
實際上:
S t r a t e g y k ( t + 1 ) = h ( E x p o s u r e k ( t ) , R e v e n u e k ( t ) , A l g o r i t h m S i g n a l s t , C o m p e t i t o r B e h a v i o r t ) . Strategy_k(t+1)
=
h(
Exposure_k(t),
Revenue_k(t),
AlgorithmSignals_t,
CompetitorBehavior_t
). S t r a t e g y k ( t + 1 ) = h ( E x p os u r e k ( t ) , R e v e n u e k ( t ) , A l g or i t hm S i g na l s t , C o m p e t i t or B e ha v i o r t ) .
因此 recommender policy 會影響:
W h a t G e t s C r e a t e d . WhatGetsCreated. W ha tG e t s C r e a t e d .
這與只影響:
W h a t G e t s S e e n WhatGetsSeen W ha tG e t s S ee n
不同。
長期 recommendation objective 應包含:
F u t u r e S u p p l y R e s p o n s e . \boxed{
FutureSupplyResponse.
} F u t u r e S u ppl y R es p o n se .
17. Algorithmic Selection Pressure
如果某類內容形式:
f 1 f_1 f 1
穩定獲得較多曝光,而:
f 2 f_2 f 2
即使品質不低也較難被推薦,
creator 會逐步改變:
P ( P r o d u c e ( f 1 ) ) ↑ . P(
Produce(f_1)
)
\uparrow. P ( P r o d u ce ( f 1 )) ↑ .
因此:
R e c o m m e n d a t i o n P o l i c y → C r e a t o r S t r a t e g y → C o n t e n t S u p p l y . RecommendationPolicy
\rightarrow
CreatorStrategy
\rightarrow
ContentSupply. R eco mm e n d a t i o n P o l i cy → C r e a t or S t r a t e g y → C o n t e n tS u ppl y .
這是一種 algorithmic selection pressure。
若 selection pressure 過強,可能得到:
C o n t e n t C o n v e r g e n c e . ContentConvergence. C o n t e n tC o n v er g e n ce .
18. Creator Ecological Collapse
本文所稱 Creator Ecological Collapse(CEC)不是指:
A l l C r e a t o r s E x i t . AllCreatorsExit. A l l C r e a t or s E x i t .
而是以下動態 regime:
cold-start observability 長期下降;
newcomer evidence accumulation 下降;
small / mid creators 的 viability rate 下降;
creator exit 或 strategic convergence 上升;
supply concentration 上升;
recommender 進一步依賴已具有大量 evidence 的 head creators;
concentration 再被強化。
形式:
E x p o s u r e C o n c e n t r a t i o n → E v i d e n c e C o n c e n t r a t i o n → C r e a t o r V i a b i l i t y G a p → S u p p l y C o n c e n t r a t i o n → E x p o s u r e C o n c e n t r a t i o n . \boxed{
ExposureConcentration
\rightarrow
EvidenceConcentration
\rightarrow
CreatorViabilityGap
\rightarrow
SupplyConcentration
\rightarrow
ExposureConcentration.
} E x p os u r e C o n ce n t r a t i o n → E v i d e n ce C o n ce n t r a t i o n → C r e a t or V iabi l i t y G a p → S u ppl y C o n ce n t r a t i o n → E x p os u r e C o n ce n t r a t i o n .
若上述閉環持續強化,則:
C E C ( t ) ↑ . CEC(t)\uparrow. C E C ( t ) ↑ .
19. Supply Concentration Feedback
令 creator supply share:
s k s u p p l y = O u t p u t k ∑ j O u t p u t j . s_k^{supply}
=
\frac{
Output_k
}{
\sum_j Output_j
}. s k s u ppl y = ∑ j O u tp u t j O u tp u t k .
供給集中度:
H H I S = ∑ k ( s k s u p p l y ) 2 . HHI_S
=
\sum_k
(s_k^{supply})^2. H H I S = k ∑ ( s k s u ppl y ) 2 .
若:
H H I S ( t + 1 ) > H H I S ( t ) HHI_S(t+1)>HHI_S(t) H H I S ( t + 1 ) > H H I S ( t )
且:
H H I E ( t + 1 ) > H H I E ( t ) , HHI_E(t+1)>HHI_E(t), H H I E ( t + 1 ) > H H I E ( t ) ,
其中 H H I E HHI_E H H I E 為 exposure concentration,
則可能存在:
S C F = S u p p l y C o n c e n t r a t i o n F e e d b a c k . \boxed{
SCF
=
Supply\ Concentration\ Feedback.
} S C F = S u ppl y C o n ce n t r a t i o n F ee d ba c k .
20. Paid Promotion as a Cold-Start Bridge
paid promotion 可以提供:
X k p a i d > 0. X_k^{paid}>0. X k p ai d > 0.
因此:
X k t o t a l = X k o r g + X k p a i d . X_k^{total}
=
X_k^{org}
+
X_k^{paid}. X k t o t a l = X k or g + X k p ai d .
在合理情況下,paid promotion 可以:
C o l d S t a r t → I n i t i a l E v i d e n c e → O r g a n i c M a t c h i n g . ColdStart
\rightarrow
InitialEvidence
\rightarrow
OrganicMatching. C o l d S t a r t → I ni t ia l E v i d e n ce → O r g ani c M a t c hin g .
這是一個 legitimate bridge。
問題不是:
P a i d P r o m o t i o n E x i s t s . PaidPromotionExists. P ai d P r o m o t i o n E x i s t s .
而是:
C a n P a i d E x p o s u r e C o n v e r t I n t o O r g a n i c V i a b i l i t y ? \boxed{
CanPaidExposureConvertIntoOrganicViability?
} C an P ai d E x p os u r e C o n v er t I n t o O r g ani c V iabi l i t y ?
21. Organic-to-Paid Substitution Risk
定義 paid exposure share:
ρ k = X k p a i d X k t o t a l + ϵ . \rho_k
=
\frac{
X_k^{paid}
}{
X_k^{total}+\epsilon
}. ρ k = X k t o t a l + ϵ X k p ai d .
若:
ρ k ↑ \rho_k\uparrow ρ k ↑
但:
X k o r g X_k^{org} X k or g
沒有隨 evidence accumulation 上升,
則 creator 對 paid exposure 形成依賴。
定義:
O P S R = P ( X t + 1 o r g ≤ X t o r g ∣ X t p a i d > 0 , E v i d e n c e t ↑ ) . OPSR
=
P(
X^{org}_{t+1}\leq X^{org}_t
\mid
X^{paid}_t>0,
Evidence_t\uparrow
). O P S R = P ( X t + 1 or g ≤ X t or g ∣ X t p ai d > 0 , E v i d e n c e t ↑ ) .
若:
O P S R ↑ , OPSR\uparrow, O P S R ↑ ,
表示 paid promotion 沒有有效成為 organic bridge,而更接近持續購買 visibility。
22. Promotion ROI
creator 付費推廣成本:
C p a i d . C_{paid}. C p ai d .
直接產生的收益:
R d i r e c t . R_{direct}. R d i r ec t .
長期 audience value:
V a u d i e n c e . V_{audience}. V a u d i e n ce .
則:
R O I p a i d = R d i r e c t + V a u d i e n c e − C p a i d C p a i d + ϵ . ROI_{paid}
=
\frac{
R_{direct}
+
V_{audience}
-
C_{paid}
}{
C_{paid}+\epsilon
}. R O I p ai d = C p ai d + ϵ R d i r ec t + V a u d i e n ce − C p ai d .
如果:
R O I p a i d < 0 ROI_{paid}<0 R O I p ai d < 0
且:
O r g a n i c R e a c h ≈ 0 , OrganicReach\approx0, O r g ani c R e a c h ≈ 0 ,
creator 面臨雙重不可行:
O r g a n i c D o e s N o t W o r k ∧ P a i d D o e s N o t W o r k . \boxed{
OrganicDoesNotWork
\land
PaidDoesNotWork.
} O r g ani cD oes N o t W or k ∧ P ai d D oes N o t W or k .
這會顯著提高:
X P k . XP_k. X P k .
23. Paid Visibility and Creator Strategy
近期 UGC platform research 顯示 paid visibility share 改變時,creators 會重新配置:
C o n t e n t E f f o r t , P a i d P r o m o t i o n , E x t e r n a l A c q u i s i t i o n . ContentEffort,
PaidPromotion,
ExternalAcquisition. C o n t e n tE f f or t , P ai d P r o m o t i o n , E x t er na l A c q u i s i t i o n .
因此平台不能只計算:
R e v e n u e F r o m P r o m o t i o n . RevenueFromPromotion. R e v e n u e F r o m P r o m o t i o n .
還應考慮:
∂ C o n t e n t E f f o r t ∂ P a i d V i s i b i l i t y , \frac{
\partial ContentEffort
}{
\partial PaidVisibility
}, ∂ P ai d V i s ibi l i t y ∂ C o n t e n tE f f or t ,
∂ A u d i e n c e G r o w t h ∂ P a i d V i s i b i l i t y , \frac{
\partial AudienceGrowth
}{
\partial PaidVisibility
}, ∂ P ai d V i s ibi l i t y ∂ A u d i e n ce G r o w t h ,
以及:
∂ E x p o s u r e I n e q u a l i t y ∂ P a i d V i s i b i l i t y . \frac{
\partial ExposureInequality
}{
\partial PaidVisibility
}. ∂ P ai d V i s ibi l i t y ∂ E x p os u r e I n e q u a l i t y .
這使 paid visibility 成為 ecosystem policy,而不只是 ad product。
24. Head Advantage Amplification
head creators 通常具有:
H i s t o r i c a l E v i d e n c e ↑ , HistoricalEvidence\uparrow, H i s t or i c a l E v i d e n ce ↑ ,
F o l l o w e r B a s e ↑ , FollowerBase\uparrow, F o l l o w er B a se ↑ ,
M o n e t i z a t i o n A b i l i t y ↑ . MonetizationAbility\uparrow. M o n e t i z a t i o n A bi l i t y ↑ .
若平台又讓:
P a i d V i s i b i l i t y PaidVisibility P ai d V i s ibi l i t y
具有強放大效果,
高資源 creator 可能同時擁有:
O r g a n i c A d v a n t a g e + P a i d A d v a n t a g e . OrganicAdvantage
+
PaidAdvantage. O r g ani c A d v an t a g e + P ai d A d v an t a g e .
此時:
E x p o s u r e G a p t + 1 > E x p o s u r e G a p t . ExposureGap_{t+1}
>
ExposureGap_t. E x p os u r e G a p t + 1 > E x p os u r e G a p t .
因此付費系統需要監測:
C a p i t a l A m p l i f i c a t i o n . \boxed{
CapitalAmplification.
} C a p i t a l A m pl i f i c a t i o n .
25. Creator Exit as an Unobserved Loss
推薦系統通常觀察:
E x i s t i n g C r e a t o r s . ExistingCreators. E x i s t in g C r e a t or s .
但離開平台的人:
k ∉ K t + 1 k\notin K_{t+1} k ∈ / K t + 1
之後不再產生內容。
因此模型容易只看到:
留下來的人表現很好。
卻忽略:
S u r v i v o r s h i p B i a s . SurvivorshipBias. S u r v i v or s hi pB ia s .
真正 ecosystem analysis 需要:
E n t r y R a t e , E x i t R a t e , D o r m a n c y R a t e , R e t u r n R a t e . EntryRate,
ExitRate,
DormancyRate,
ReturnRate. E n t r y R a t e , E x i tR a t e , D or man cy R a t e , R e t u r n R a t e .
26. Creator Cohort Survival
對第 τ \tau τ 期加入的 creator cohort:
K τ . K_\tau. K τ .
定義 survival:
S τ ( Δ ) = ∣ { k ∈ K τ : A c t i v e ( k , τ + Δ ) = 1 } ∣ ∣ K τ ∣ . S_\tau(\Delta)
=
\frac{
|\{
k\in K_\tau:
Active(k,\tau+\Delta)=1
\}|
}{
|K_\tau|
}. S τ ( Δ ) = ∣ K τ ∣ ∣ { k ∈ K τ : A c t i v e ( k , τ + Δ ) = 1 } ∣ .
可以比較:
S τ 1 , S τ 2 S_{\tau_1},
S_{\tau_2} S τ 1 , S τ 2
在推薦政策改版前後的差異。
若:
S n e w ( Δ ) ↓ S_{new}(\Delta)\downarrow S n e w ( Δ ) ↓
同時:
H e a d E x p o s u r e S h a r e ↑ , HeadExposureShare\uparrow, H e a d E x p os u r e S ha r e ↑ ,
則值得進一步檢查 causal relation。
27. Time-to-First-Sustainable-Audience
定義 sustainable audience threshold:
A ∗ . A^*. A ∗ .
對 creator k k k :
T T S A k = inf { t : A k ( t ) ≥ A ∗ } . TTSA_k
=
\inf
\{
t:
A_k(t)\geq A^*
\}. T T S A k = inf { t : A k ( t ) ≥ A ∗ } .
如果:
T T S A ↑ TTSA\uparrow T T S A ↑
表示 newcomers 需要更長時間才能建立基本 audience stock。
若:
T T S A → ∞ , TTSA\rightarrow\infty, T T S A → ∞ ,
則該 creator 在 observation window 內沒有形成可持續 audience。
28. Cold-Start Test Completion Rate
不是所有新人都需要成功。
但平台至少應讓一定比例完成:
initial evidence acquisition . \text{initial evidence acquisition}. initial evidence acquisition .
令最低 evidence threshold:
E ∗ . E^*. E ∗ .
定義:
C S T R = P ( E E k ≥ E ∗ ∣ k ∈ K n e w ) . CSTR
=
P(
EE_k\geq E^*
\mid
k\in K_{new}
). C S T R = P ( E E k ≥ E ∗ ∣ k ∈ K n e w ) .
其中:
E E k EE_k E E k
是 Effective Evidence。
若:
C S T R ↓ , CSTR\downarrow, C S T R ↓ ,
表示更多新人連「被系統有效測試」都沒有完成。
29. Creator Viability Rate
定義:
C V R = P ( C E R k ≥ C V T k ) . CVR
=
P(
CER_k\geq CVT_k
). C V R = P ( C E R k ≥ C V T k ) .
分 cohort:
C V R n e w , C V R s m a l l , C V R m i d , C V R h e a d . CVR_{new},
CVR_{small},
CVR_{mid},
CVR_{head}. C V R n e w , C V R s ma l l , C V R mi d , C V R h e a d .
平台應監測:
G a p C V R = C V R h e a d − C V R n e w . Gap_{CVR}
=
CVR_{head}
-
CVR_{new}. G a p C V R = C V R h e a d − C V R n e w .
gap 本身不必為零。
但若:
G a p C V R ↑ Gap_{CVR}\uparrow G a p C V R ↑
且:
C S O ↓ , C S T R ↓ , CSO\downarrow,
CSTR\downarrow, C S O ↓ , C S T R ↓ ,
則供給側風險上升。
30. Creator Diversity Retention
令 active creators 的 semantic / topical clusters:
G K . \mathcal{G}_K. G K .
定義:
C D R ( T ) = D C ( t + T ) D C ( t ) + ϵ . CDR(T)
=
\frac{
D_C(t+T)
}{
D_C(t)+\epsilon
}. C D R ( T ) = D C ( t ) + ϵ D C ( t + T ) .
若:
C D R < 1 CDR<1 C D R < 1
長期且持續下降,
表示 active creator diversity 收縮。
可以再區分:
D c r e a t o r , D t o p i c , D f o r m a t , D v i e w p o i n t . D_{creator},
D_{topic},
D_{format},
D_{viewpoint}. D cr e a t or , D t o p i c , D f or ma t , D v i e w p o in t .
31. Platform Utility with Future Supply
短期平台 objective:
U t = R e v e n u e t + E n g a g e m e n t t . U_t
=
Revenue_t
+
Engagement_t. U t = R e v e n u e t + E n g a g e m e n t t .
更完整的長期 objective:
U L T = ∑ t = 0 T γ t [ U u s e r , t + α U c r e a t o r , t + β D s u p p l y , t + η R e v e n u e t ] . U^{LT}
=
\sum_{t=0}^{T}
\gamma^t
[
U_{user,t}
+
\alpha U_{creator,t}
+
\beta D_{supply,t}
+
\eta Revenue_t
]. U L T = t = 0 ∑ T γ t [ U u ser , t + α U cr e a t or , t + β D s u ppl y , t + η R e v e n u e t ] .
其中:
D s u p p l y , t D_{supply,t} D s u ppl y , t
代表未來內容供給的健康度。
若:
γ \gamma γ
過低,
平台可能選擇:
S h o r t T e r m E x t r a c t i o n ShortTermExtraction S h or tT er m E x t r a c t i o n
而忽略:
C r e a t o r R e t e n t i o n . CreatorRetention. C r e a t or R e t e n t i o n .
32. Ecosystem Health Metrics
本文建議至少監測:
32.1 Cold-Start Observability
C S O . CSO. C S O .
32.2 Cold-Start Test Completion Rate
C S T R . CSTR. C S T R .
32.3 Time-to-First-Sustainable-Audience
T T S A . TTSA. T T S A .
32.4 Creator Viability Rate
C V R . CVR. C V R .
32.5 Creator Exit Rate
C E R e x i t = N e x i t N a c t i v e . CER_{exit}
=
\frac{
N_{exit}
}{
N_{active}
}. C E R e x i t = N a c t i v e N e x i t .
32.6 Organic-to-Paid Substitution Risk
O P S R . OPSR. O P S R .
32.7 Supply HHI
H H I S . HHI_S. H H I S .
32.8 Creator Diversity Retention
C D R . CDR. C D R .
33. A Minimal Sustainable Cold-Start Architecture
本文提出一個最小架構。
Stage 1 — Eligibility
新內容先通過:
S a f e t y ∩ S p a m ∩ M i n i m u m Q u a l i t y ∩ B a s i c R e l e v a n c e . Safety
\cap
Spam
\cap
MinimumQuality
\cap
BasicRelevance. S a f e t y ∩ S p am ∩ M inim u m Q u a l i t y ∩ B a s i c R e l e v an ce .
Stage 2 — Exploration Budget
分配:
X m i n . X_{min}. X min .
Stage 3 — Evidence Acquisition
收集:
A c t i v e E n g a g e m e n t , S a v e , F o l l o w , S e a r c h R e e n t r y , C r e a t o r V i s i t . ActiveEngagement,
Save,
Follow,
SearchReentry,
CreatorVisit. A c t i v e E n g a g e m e n t , S a v e , F o l l o w , S e a r c h R ee n t r y , C r e a t or V i s i t .
Stage 4 — Audience Matching
建立:
P ( P o s i t i v e E n g a g e m e n t ∣ U s e r C l u s t e r ) . P(
PositiveEngagement
\mid
UserCluster
). P ( P os i t i v e E n g a g e m e n t ∣ U ser C l u s t er ) .
Stage 5 — Organic Expansion
若 evidence positive:
E x p o s u r e ↑ . Exposure\uparrow. E x p os u r e ↑ .
Stage 6 — Controlled Decay
若有足夠 exposure 但 evidence consistently poor:
E x p o s u r e ↓ . Exposure\downarrow. E x p os u r e ↓ .
這裡關鍵是:
D o N o t D e c a y B e f o r e T e s t i n g . \boxed{
DoNotDecayBeforeTesting.
} D o N o t D ec a y B e f or e T es t in g .
34. Paid Promotion Boundary
paid promotion 應與 cold-start test 分離。
最低 organic testing opportunity:
X m i n o r g X_{min}^{org} X min or g
不應完全依賴:
P a y m e n t . Payment. P a y m e n t .
否則:
A b i l i t y T o P a y AbilityToPay A bi l i t y T o P a y
會替代:
A b i l i t y T o A t t r a c t A u d i e n c e . AbilityToAttractAudience. A bi l i t y T o A tt r a c t A u d i e n ce .
平台可以提供 paid amplification,但應保留:
O r g a n i c T e s t a b i l i t y . \boxed{
OrganicTestability.
} O r g ani c T es t abi l i t y .
35. Creator Adaptation Audit
推薦系統改版後,不應只觀察:
C T R , W a t c h T i m e . CTR,
WatchTime. C T R , W a t c h T im e .
也應追蹤 creators 是否:
更新頻率改變;
production cost 改變;
topic distribution 收斂;
thumbnail / title strategy 收斂;
paid promotion 上升;
cross-platform publishing 上升;
exit 上升。
這些都是:
P o l i c y → S u p p l y R e s p o n s e Policy
\rightarrow
SupplyResponse P o l i cy → S u ppl y R es p o n se
的證據。
36. Empirical Protocol
36.1 Creator cohort design
建立:
K n e w , K s m a l l , K m i d , K h e a d . K_{new},
K_{small},
K_{mid},
K_{head}. K n e w , K s ma l l , K mi d , K h e a d .
36.2 Exposure tracking
記錄:
X k h o m e , X k f o l l o w , X k s e a r c h , X k p a i d . X_k^{home},
X_k^{follow},
X_k^{search},
X_k^{paid}. X k h o m e , X k f o l l o w , X k se a r c h , X k p ai d .
36.3 Evidence tracking
記錄:
E E k . EE_k. E E k .
36.4 Audience accumulation
估計:
A k ( t ) . A_k(t). A k ( t ) .
36.5 Creator persistence
計算:
S τ ( Δ ) . S_\tau(\Delta). S τ ( Δ ) .
36.6 Version comparison
比較推薦改版前後:
Δ C S O , Δ C S T R , Δ T T S A , Δ C V R , Δ E x i t R a t e , Δ H H I S . \Delta CSO,
\Delta CSTR,
\Delta TTSA,
\Delta CVR,
\Delta ExitRate,
\Delta HHI_S. Δ C S O , Δ C S T R , Δ T T S A , Δ C V R , Δ E x i tR a t e , Δ H H I S .
37. Causal Boundaries
即使觀察到:
C r e a t o r E x i t ↑ CreatorExit\uparrow C r e a t or E x i t ↑
與:
E x p o s u r e C o n c e n t r a t i o n ↑ , ExposureConcentration\uparrow, E x p os u r e C o n ce n t r a t i o n ↑ ,
也不能直接推出:
E x p o s u r e C o n c e n t r a t i o n → C r e a t o r E x i t . ExposureConcentration
\rightarrow
CreatorExit. E x p os u r e C o n ce n t r a t i o n → C r e a t or E x i t .
creator exit 還可能受到:
經濟景氣;
competitor platform;
creator life cycle;
monetization policy;
content regulation;
production technology;
audience trend。
因此需要:
staggered policy changes;
difference-in-differences;
matched creator cohorts;
randomized traffic experiments;
regression discontinuity;
synthetic controls;
creator surveys。
本文提出的是:
M e c h a n i s m H y p o t h e s i s \boxed{
MechanismHypothesis
} M ec hani s m H y p o t h es i s
與可檢驗 metrics,不是對單一平台做無證據的因果斷言。
38. Creator Ecological Collapse as a Regime, Not a Binary Event
CEC 應被理解為:
C E C I n d e x ( t ) ∈ [ 0 , 1 ] . CECIndex(t)\in[0,1]. C E C I n d e x ( t ) ∈ [ 0 , 1 ] .
例如:
C E C I n d e x = w 1 H H I ~ E + w 2 H H I ~ S + w 3 ( 1 − C S T R ) + w 4 ( 1 − C V R n e w ) + w 5 E x i t R a t e + w 6 O P S R . CECIndex
=
w_1\tilde{HHI}_E
+
w_2\tilde{HHI}_S
+
w_3(1-CSTR)
+
w_4(1-CVR_{new})
+
w_5ExitRate
+
w_6OPSR. C E C I n d e x = w 1 H H I ~ E + w 2 H H I ~ S + w 3 ( 1 − C S T R ) + w 4 ( 1 − C V R n e w ) + w 5 E x i tR a t e + w 6 O P S R .
它不是:
P l a t f o r m A l i v e ↔ P l a t f o r m D e a d . PlatformAlive
\leftrightarrow
PlatformDead. P l a t f or m A l i v e ↔ P l a t f or m D e a d .
而是一個供給側退化程度。
這讓研究者可以觀察:
d C E C I n d e x d t . \frac{dCECIndex}{dt}. d t d C E C I n d e x .
39. Relation to A06
A05 已經產生一個重要矛盾。
平台可能看到:
C T R ↑ , W a t c h T i m e ↑ , A d R e v e n u e ↑ . CTR\uparrow,
WatchTime\uparrow,
AdRevenue\uparrow. C T R ↑ , W a t c h T im e ↑ , A d R e v e n u e ↑ .
同時:
C S O ↓ , C S T R ↓ , C V R n e w ↓ , E x i t R a t e ↑ . CSO\downarrow,
CSTR\downarrow,
CVR_{new}\downarrow,
ExitRate\uparrow. C S O ↓ , C S T R ↓ , C V R n e w ↓ , E x i tR a t e ↑ .
因此:
S h o r t T e r m M e t r i c s I m p r o v e ∧ L o n g T e r m E c o s y s t e m D e g r a d e s . \boxed{
ShortTermMetricsImprove
\land
LongTermEcosystemDegrades.
} S h or tT er m M e t r i cs I m p r o v e ∧ L o n g T er m E cosy s t e m D e g r a d es .
完全可能同時成立。
A06 將正式處理這個問題:
M e t r i c S u c c e s s ≠ P r o d u c t S u c c e s s . \boxed{
MetricSuccess
\neq
ProductSuccess.
} M e t r i c S u ccess = P r o d u c tS u ccess .
40. Limitations
第一,creator motivation 高度異質,並非所有 creators 都以收入最大化為目標,因此 C E R CER C E R 與 C V T CVT C V T 應依 cohort 或個體校準。
第二,公平曝光不等於平均曝光。低品質、spam、違規或高度不相關內容不應僅因 cold start 而獲得無限制流量。
第三,creator exit 具有多重原因,推薦政策只是其中一個可能因子。
第四,付費推廣不是本質有害。合理 paid promotion 可以作為 audience discovery accelerator。風險在於 organic testability 消失或 paid dependence 長期上升。
第五,本文的 CEC 是平台生態模型,不主張任何特定現實平台已達到 ecological collapse。
41. Conclusion
本文將 cold start 從單純預測問題提升為 creator entry 與平台供給問題。
最基本的鏈條是:
E x p o s u r e → E v i d e n c e → A u d i e n c e M a t c h i n g → A u d i e n c e S t o c k → C r e a t o r R e t u r n . \boxed{
Exposure
\rightarrow
Evidence
\rightarrow
AudienceMatching
\rightarrow
AudienceStock
\rightarrow
CreatorReturn.
} E x p os u r e → E v i d e n ce → A u d i e n ce M a t c hin g → A u d i e n ce S t oc k → C r e a t or R e t u r n .
因此:
N o E x p o s u r e ≠ N e g a t i v e D e m a n d E v i d e n c e \boxed{
NoExposure
\neq
NegativeDemandEvidence
} N o E x p os u r e = N e g a t i v eD e man d E v i d e n ce
以及:
N o E v i d e n c e ≠ L o w Q u a l i t y . \boxed{
NoEvidence
\neq
LowQuality.
} N o E v i d e n ce = L o w Q u a l i t y .
若推薦系統要求:
E v i d e n c e Evidence E v i d e n ce
才能給:
E x p o s u r e , Exposure, E x p os u r e ,
但又要求:
E x p o s u r e Exposure E x p os u r e
才能生成:
E v i d e n c e , Evidence, E v i d e n ce ,
則 newcomer 會陷入:
E v i d e n c e - S t a r v e d C o l d S t a r t . \boxed{
Evidence\text{-}Starved\ Cold\ Start.
} E v i d e n ce - S t a r v e d C o l d S t a r t .
平台應至少保留:
O r g a n i c T e s t a b i l i t y \boxed{
OrganicTestability
} O r g ani c T es t abi l i t y
使通過最低 eligibility gate 的新人具有:
P ( B e i n g M e a n i n g f u l l y T e s t e d ) > 0. P(
BeingMeaningfullyTested
)>0. P ( B e in g M e anin g f u l l y T es t e d ) > 0.
本文並提出:
C V T , O D V , E E C , C E R , X P , C S T R , C V R , O P S R , S C F CVT,
ODV,
EEC,
CER,
XP,
CSTR,
CVR,
OPSR,
SCF C V T , O D V , E E C , C E R , X P , C S T R , C V R , O P S R , S C F
等概念,將 creator survival 與推薦曝光放在同一套動態模型中。
最終,Creator Ecological Collapse 被定義為一種可能的自我強化 regime:
E x p o s u r e C o n c e n t r a t i o n → E v i d e n c e C o n c e n t r a t i o n → C r e a t o r V i a b i l i t y G a p → S u p p l y C o n c e n t r a t i o n → E x p o s u r e C o n c e n t r a t i o n . \boxed{
ExposureConcentration
\rightarrow
EvidenceConcentration
\rightarrow
CreatorViabilityGap
\rightarrow
SupplyConcentration
\rightarrow
ExposureConcentration.
} E x p os u r e C o n ce n t r a t i o n → E v i d e n ce C o n ce n t r a t i o n → C r e a t or V iabi l i t y G a p → S u ppl y C o n ce n t r a t i o n → E x p os u r e C o n ce n t r a t i o n .
這並不是說平台必然走向崩潰,而是指出:
R e c o m m e n d a t i o n P o l i c y shapes not only what users see, but also what creators will continue to create. \boxed{
RecommendationPolicy
\text{ shapes not only what users see, but also what creators will continue to create.}
} R eco mm e n d a t i o n P o l i cy shapes not only what users see, but also what creators will continue to create.
因此推薦系統的長期 objective 不能只最大化今天的 consumption。
它必須同時考慮:
F u t u r e C o n t e n t S u p p l y . \boxed{
FutureContentSupply.
} F u t u r e C o n t e n tS u ppl y .
至此,A01–A05 形成:
O b s e r v a t i o n → P r e f e r e n c e I n t e r p r e t a t i o n → E n d o g e n o u s C o n t a m i n a t i o n → E x p o s u r e C o n c e n t r a t i o n → S u p p l y R e s p o n s e . \boxed{
Observation
\rightarrow
PreferenceInterpretation
\rightarrow
EndogenousContamination
\rightarrow
ExposureConcentration
\rightarrow
SupplyResponse.
} O b ser v a t i o n → P r e f er e n ce I n t er p r e t a t i o n → E n d o g e n o u s C o n t amina t i o n → E x p os u r e C o n ce n t r a t i o n → S u ppl y R es p o n se .
下一篇 A06 將收束整個 Series A,處理平台為何可能在 CTR、watch time、廣告收入與推薦命中率都變好的同時,逐漸失去使用者效用、創作者供給與長期平台價值。
References
[0] Neo.K / EveMissLab. “Recommendation as an Observation Operator.” Series A, Paper A01, v0.1, 2026.
[1] Neo.K / EveMissLab. “The Platform-Induced Exposure Bubble.” Series A, Paper A04, v0.1, 2026.
[2] Liu, J.-H., Zhou, T., Zhang, Z.-K., Yang, Z., Liu, C., & Li, W.-M. “Promoting Cold-Start Items in Recommender Systems.” PLOS ONE, 9(12), e113457, 2014. DOI: 10.1371/journal.pone.0113457.
[3] Zhu, Z., Kim, J., Nguyen, T., Fenton, A., & Caverlee, J. “Fairness among New Items in Cold Start Recommender Systems.” Proceedings of the 44th International ACM SIGIR Conference on Research and Development in Information Retrieval, 2021.
[4] Gómez, E., Boratto, L., & Salamó, M. “Provider Fairness Across Continents in Collaborative Recommender Systems.” Information Processing & Management, 59(1), 102719, 2022. DOI: 10.1016/j.ipm.2021.102719.
[5] Patro, G. K., Biswas, A., Ganguly, N., Gummadi, K. P., & Chakraborty, A. “FairRec: Two-Sided Fairness for Personalized Recommendations in Two-Sided Platforms.” Proceedings of The Web Conference 2020, 2020. arXiv:2002.10764.
[6] Wu, Y., Cao, J., Xu, G., & Tan, Y. “TFROM: A Two-sided Fairness-Aware Recommendation Model for Both Customers and Providers.” Proceedings of the 44th International ACM SIGIR Conference on Research and Development in Information Retrieval, pp. 1013–1022, 2021. DOI: 10.1145/3404835.3462882.
[7] Sun, Y., & Sun, B. “From Exposure to Followers: A Stock-and-Flow Closed-Loop Framework of Creator Dynamics.” Information Processing & Management, 63(5), 104677, 2026. DOI: 10.1016/j.ipm.2026.104677.
[8] Hron, J., Krauth, K., Jordan, M. I., Kilbertus, N., & Dean, S. “Modeling Content Creator Incentives on Algorithm-Curated Platforms.” arXiv:2206.13102, 2022.
[9] Zhao, W., Feng, H., Feng, N., & Li, M. “Does Paying for Visibility Pay Off? The Impact of Sponsored Recommendations on UGC Platforms.” Decision Support Systems, 206, 114668, 2026. DOI: 10.1016/j.dss.2026.114668.
[10] Montero-Porras, E., Smets, A., & Lenaerts, T. “Enshittification of Algorithmic Recommendation: An Evolutionary Model of Platform-Creator Dynamics.” HHAI 2026, 2026. DOI: 10.3233/FAIA260512.
[11] Hu, R., Feng, H., Tayi, G. K., & Feng, N. “Reward Strategies for Content Platforms: Equity or Equality?” Information & Management, 63(4), 2026.
Series Continuation
A06 — Metric Success, Product Failure