← Archive
lm-003337 · 2026-09

指標成功與產品失敗:推薦平台的代理目標錯位、多方效用與長期生態治理

下載 MD 檔 ⬇

Series A — Algorithmic Observation, Recommendation & Platform Ecology

Paper A06 — Metric Success, Product Failure

指標成功與產品失敗:推薦平台的代理目標錯位、多方效用與長期生態治理

English Title: Metric Success, Product Failure: Proxy Objective Misalignment, Multi-Stakeholder Utility, and Long-Term Ecosystem Governance in Recommender Platforms
Series: Algorithmic Observation, Recommendation & Platform Ecology
Paper ID: A06
Version: v0.1
Date: 2026-08-31
Status: Canonical UTF-8 Source
Author: Neo.K / EveMissLab


Abstract

現代推薦平台高度依賴可量化指標,例如 click-through rate、watch time、completion rate、return frequency、advertising revenue、conversion、retention 與 recommendation hit rate。這些指標具有工程可測量性與商業可操作性,因此自然成為模型優化、產品實驗與組織績效的核心。然而,平台真正關心的價值通常比任何單一 metric 更廣:使用者是否得到有價值的資訊與娛樂、是否願意長期回訪、創作者是否存在可持續生產的回報、內容供給是否保持多樣、廣告主是否獲得真實有效注意力,以及平台本身是否維持長期商業與社群資產。

本文延續 Series A 前五篇,提出 Metric–Product Separation framework。核心命題是:

MetricSuccessProductSuccess.\boxed{ MetricSuccess \neq ProductSuccess. }

本文將推薦平台視為 multi-stakeholder dynamic system,定義使用者效用、創作者效用、廣告主效用、平台短期收益、平台長期價值與內容供給健康度。本文分析 proxy optimization 如何經由 Goodhart-style failure、selection effects、survivorship bias、endogenous telemetry、exposure concentration 與 delayed externalities,使局部 KPI 上升但整體產品價值下降。

本文提出 Local Metric Improvement / Global Utility Degradation(LMIGUD)、Proxy Capture Ratio(PCR)、Delayed Ecosystem Cost(DEC)、Metric-Induced Behavior Shift(MIBS)、Stakeholder Utility Divergence(SUD)、Product Health Gap(PHG)與 Long-Horizon Platform Utility(LHPU)等概念。本文進一步區分 leading metrics、lagging metrics 與 hidden-state metrics,指出 creator exit、content supply concentration、user agency loss 與 advertiser attention quality 常比財務收入更晚或更難顯現。

本文最後提出一個多層治理架構:短期 ranking objective 不直接等同 platform objective;A/B test 不應只看 engagement uplift,而應配套 guardrail metrics、cohort retention、creator viability、long-tail observability、explicit-preference violation、paid-promotion conversion、surveyed satisfaction 與 delayed ecosystem audit。本文並將 Home、Following、Explore、Popular、Search 等觀察 surface 視為不同產品契約,以降低單一 proxy 對全部資訊入口的控制。

Series A 至此完成一個完整閉環:推薦系統決定可觀察世界、推論使用者偏好、由自身曝光污染行為證據、集中平台曝光、改變創作者供給,最後又可能因局部 KPI 成功而錯誤判定系統健康。本文的核心結論是:推薦平台真正需要優化的不是「更多互動」,而是多方、長期、可持續的資訊生態效用。

Keywords: recommender systems; Goodhart's law; proxy metrics; multi-stakeholder recommendation; user satisfaction; creator economy; platform governance; engagement; long-term utility; ecosystem health


1. Introduction

推薦系統工程非常擅長回答:

DidCTRIncrease?DidCTRIncrease? DidWatchTimeIncrease?DidWatchTimeIncrease? DidConversionIncrease?DidConversionIncrease? DidRevenueIncrease?DidRevenueIncrease?

這些問題重要,而且不能被忽略。

但平台真正需要回答的問題通常是:

DidTheProductBecomeBetter?\boxed{ DidTheProductBecomeBetter? }

兩者不是同一問題。

一個推薦改版可能:

CTR,CTR\uparrow, WatchTime,WatchTime\uparrow, Views,Views\uparrow,

甚至:

Revenue.Revenue\uparrow.

同時也可能:

UserAgency,UserAgency\downarrow, LongTailObservability,LongTailObservability\downarrow, CreatorEntryViability,CreatorEntryViability\downarrow, InformationDiversity.InformationDiversity\downarrow.

這種情況並不矛盾。

因為前一組量:

MM

是可觀測 proxy metrics,

後一組量:

HH

是更廣的 product / ecosystem health state。

若:

M≢H,M \not\equiv H,

則:

maxM\max M

不保證:

maxH.\max H.

本文因此提出 Series A 的收束命題:

MetricSuccessProductSuccess.\boxed{ MetricSuccess \neq ProductSuccess. }

2. Series A Recap

Series A 前五篇建立:

A01 — Observation

CtOu,t(s)Lu,t(s).\mathcal{C}_t \xrightarrow{ \mathcal{O}_{u,t}^{(s)} } L_{u,t}^{(s)}.

推薦是 visibility allocation 與 observation policy。

A02 — Preference Semantics

AwarenessInterestIntentFeasibility.Awareness \neq Interest \neq Intent \neq Feasibility.

A03 — Endogenous Preference Contamination

RecommendationExposureTelemetryPreferenceEstimate.Recommendation \rightarrow Exposure \rightarrow Telemetry \rightarrow PreferenceEstimate.

平台可能用自己產生的行為資料證明自己的推論。

A04 — Platform-Induced Exposure Bubble

TopicDiversityCreatorDiversityExposureDiversity.TopicDiversity \neq CreatorDiversity \neq ExposureDiversity.

A05 — Creator Ecology

ExposureEvidenceAudienceCreatorReturnFutureSupply.Exposure \rightarrow Evidence \rightarrow Audience \rightarrow CreatorReturn \rightarrow FutureSupply.

A06 問的是:

如果以上各層正在退化,但平台核心 KPI 仍然增加,組織是否可能認為改版成功?

答案是:

Yes.\boxed{ Yes. }

而這正是最難偵測的 failure mode。


3. Related Work

3.1 Accuracy is not enough

Recommender-system evaluation 很早就指出 predictive accuracy 並不足以衡量推薦品質。McNee、Riedl 與 Konstan 的經典論文 “Being Accurate is Not Enough” 強調 novelty、serendipity、diversity 等 beyond-accuracy qualities 對實際推薦價值的重要性 [2]。

因此:

PredictionAccuracyRecommendationUtility.PredictionAccuracy \neq RecommendationUtility.

3.2 User experience

Knijnenburg 等人提出 recommender systems user-experience framework,區分客觀系統面、使用者主觀感受、互動與個人/情境特徵 [3]。

這表示即使推薦系統 offline metrics 很好:

OfflineScoreOfflineScore\uparrow

也不能直接推出:

UserExperience.UserExperience\uparrow.

3.3 Multi-stakeholder recommendation

Burke、Abdollahpouri、Mobasher 等人指出,推薦系統往往存在多個利益相關者,包括 users、providers、platforms 等,推薦價值不應只以單一 consumer utility 衡量 [4][5]。

因此平台 objective 更接近:

U=f(Uconsumer,Uprovider,Uplatform,).U = f( U_{consumer}, U_{provider}, U_{platform}, \ldots ).

3.4 Provider fairness and ecosystem effects

Two-sided fairness 與 provider fairness research 進一步指出,推薦排序決定商品或內容提供者獲得的 exposure,而只最大化 consumer-side metric 可能導致不平衡的 provider opportunities [6][7]。

3.5 Feedback loops

Series A 已引用的大量研究顯示,推薦輸出會影響後續資料、熱門度與使用者行為 [8][9]。

因此:

MetrictMetric_t

不是被動測量世界。

它可能部分來自:

PolicytPolicy_t

本身。

3.6 Goodhart-style proxy failure

Goodhart-style reasoning 可概括為:

當某個 measure 被強烈用作 target 時,它與真正目標之間原本穩定的關係可能失效。

Manheim 與 Garrabrant 將 Goodhart effects 分類為 regressional、extremal、causal 與 adversarial forms [10]。

推薦平台特別容易出現:

CausalGoodhartCausalGoodhart

因為平台直接改變:

Exposure,Behavior,Telemetry.Exposure, Behavior, Telemetry.

4. Metric–Product Separation

令:

Mt=(CTRt,WatchTimet,Revenuet,Retentiont,Conversiont)M_t = ( CTR_t, WatchTime_t, Revenue_t, Retention_t, Conversion_t )

為 observable metrics。

令:

Ht=(UserUtilityt,CreatorHealtht,SupplyDiversityt,Agencyt,AdvertiserValuet,LongTermPlatformValuet)H_t = ( UserUtility_t, CreatorHealth_t, SupplyDiversity_t, Agency_t, AdvertiserValue_t, LongTermPlatformValue_t )

為 product-health state。

Metric model:

Mt=g(Ht,Policyt,UIt,Measurementt,ExternalStatet).M_t = g( H_t, Policy_t, UI_t, Measurement_t, ExternalState_t ).

注意:

MtM_t

不是:

HtH_t

的直接同義詞。

因此:

Mt is an observation of the product, not the product itself.\boxed{ M_t \text{ is an observation of the product, not the product itself.} }

這與 A01 的 observation distinction 在產品治理層重新出現。


5. Proxy Objective

真實平台目標:

Gt.G_t.

GtG_t 通常不可直接即時測量。

因此平台使用 proxy:

Mt.M_t.

訓練或產品決策變成:

maxπE[Mtπ].\max_\pi E[ M_t \mid \pi ].

但真正希望:

maxπE[Gtπ].\max_\pi E[ G_t \mid \pi ].

只有在:

Corr(M,G)Corr(M,G)

在 intervention 後仍穩定時,

前者才可靠代表後者。

Goodhart failure 發生於:

Corr(M,Gdo(π)).Corr( M,G \mid do(\pi) ) \downarrow.

6. Causal Metric Inflation

A03 已指出:

RecommendationExposureTelemetry.Recommendation \rightarrow Exposure \rightarrow Telemetry.

因此平台可以透過 UI 或 policy 直接提升 metric。

例如:

AutoplayWatchEvents.Autoplay\uparrow \Rightarrow WatchEvents\uparrow. RepeatExposureImpressions.RepeatExposure\uparrow \Rightarrow Impressions\uparrow. HeadContentSharePredictedCTRAccuracy.HeadContentShare\uparrow \Rightarrow PredictedCTRAccuracy\uparrow.

這些數字變化未必代表:

UserPreferenceAlignment.UserPreferenceAlignment\uparrow.

本文稱:

MetricInflationcausal.\boxed{ MetricInflation_{causal}. }

7. Local Metric Improvement / Global Utility Degradation

本文定義:

LMIGUD.\boxed{ LMIGUD. }

若存在某 policy change:

π0π1\pi_0 \rightarrow \pi_1

使:

M(π1)>M(π0)M(\pi_1)>M(\pi_0)

但:

Uglobal(π1)<Uglobal(π0),U^{global}(\pi_1) < U^{global}(\pi_0),

則發生:

LMIGUD=1.LMIGUD=1.

其中:

UglobalU^{global}

可以包含多 stakeholder 與長期效用。

這是:

MetricSuccessProductFailure.\boxed{ MetricSuccess \land ProductFailure. }

的形式化定義。


8. Multi-Stakeholder Utility

令 stakeholders:

S={User,Creator,Advertiser,Platform,Society}.\mathcal{S} = \{ User, Creator, Advertiser, Platform, Society \}.

總效用:

UtMS=sSwsUs(t).U^{MS}_t = \sum_{s\in\mathcal{S}} w_s U_s(t).

其中:

wsw_s

不是自然常數,而是 platform governance choice。

例如:

Uuser=f(Relevance,Satisfaction,Agency,TimeValue,InformationGain).U_{user} = f( Relevance, Satisfaction, Agency, TimeValue, InformationGain ). Ucreator=f(Exposure,AudienceGrowth,Revenue,Predictability,Viability).U_{creator} = f( Exposure, AudienceGrowth, Revenue, Predictability, Viability ). Uadvertiser=f(QualifiedAttention,Conversion,BrandLift,Incrementality).U_{advertiser} = f( QualifiedAttention, Conversion, BrandLift, Incrementality ). Uplatform=f(Revenue,Retention,Risk,FutureSupply,Trust).U_{platform} = f( Revenue, Retention, Risk, FutureSupply, Trust ).

9. Platform Objective Is Not Ranking Objective

Ranking model 可能優化:

Jrank=αCTR+βWatchTime+γConversion.J_{rank} = \alpha CTR + \beta WatchTime + \gamma Conversion.

但 platform objective:

JplatformJ_{platform}

應該更廣。

因此:

JrankJplatform.\boxed{ J_{rank} \neq J_{platform}. }

合理架構是:

JrankJplatformJ_{rank} \subset J_{platform}

並且:

JplatformJ_{platform}

對:

JrankJ_{rank}

施加 constraints / guardrails。


10. Proxy Capture Ratio

本文提出 Proxy Capture Ratio:

PCR=ΔMΔUtarget+ϵ.PCR = \frac{ \Delta M }{ \Delta U^{target}+\epsilon }.

如果:

ΔM0\Delta M\gg0

但:

ΔUtarget0,\Delta U^{target}\approx0,

則:

PCR1.PCR\gg1.

代表 policy 很有效地提升 metric,但沒有等比例提升真正目標。

若:

ΔUtarget<0,\Delta U^{target}<0,

則 proxy capture 更嚴重。


11. Engagement Is Not Satisfaction

使用者觀看很久可能代表:

  • 真正喜歡;
  • 任務需要;
  • autoplay;
  • sunk-cost;
  • search friction;
  • interface trapping;
  • 情緒性 consumption;
  • 找不到替代內容。

因此:

WatchTimeSatisfaction.WatchTime \neq Satisfaction.

更精確:

WatchTime=f(Satisfaction,Friction,Habit,Autoplay,Availability,Intent,ContentLength).WatchTime = f( Satisfaction, Friction, Habit, Autoplay, Availability, Intent, ContentLength ).

因此:

maxWatchTime\max WatchTime

不能自然等價:

maxSatisfaction.\max Satisfaction.

12. Retention Is Not Loyalty

同樣:

RetentionRetention

可能來自:

HighUtilityHighUtility

但也可能來自:

HighSwitchingCost,NetworkEffects,ContentLockIn,LackOfAlternatives.HighSwitchingCost, NetworkEffects, ContentLockIn, LackOfAlternatives.

因此:

RetentionRetention\uparrow

不能單獨證明:

Trust.Trust\uparrow.

真正要觀察:

VoluntaryPreferenceVoluntaryPreference

與:

Dependence.Dependence.

13. Views Are Not Qualified Attention

廣告/創作者經濟常使用:

Views,Impressions,WatchTimeViews, Impressions, WatchTime

代表 attention。

但真正具有商業價值的是:

QualifiedAttention.QualifiedAttention.

定義:

QA=ActiveAttention×AudienceFit×Intentionality.QA = ActiveAttention \times AudienceFit \times Intentionality.

因此:

1,000,0001,000,000

個 autoplay-heavy views,

可能小於:

100,000100,000

個 high-intent views 的 advertiser value。


14. Advertising Metric Dilution

如果:

ViewsViews\uparrow

主要來自:

PassiveExposure,PassiveExposure\uparrow,

則:

CPVCPV

可能表面改善,

但:

ConversionPerQualifiedViewConversionPerQualifiedView

可能下降。

本文定義:

AMD=1QualifiedAttentionGrowthRawViewGrowth+ϵ.AMD = 1- \frac{ QualifiedAttentionGrowth }{ RawViewGrowth+\epsilon }.

即 Advertising Metric Dilution。

若:

AMD,AMD\uparrow,

raw traffic 與真正商業 attention 的關係正在變弱。


15. Short-Term Revenue versus Long-Term Asset

平台短期收入:

Rt.R_t.

平台長期資產:

Atplatform=(UserTrust,CreatorSupply,ContentGraph,Brand,AdvertiserConfidence,SocialRelations).A_t^{platform} = ( UserTrust, CreatorSupply, ContentGraph, Brand, AdvertiserConfidence, SocialRelations ).

長期價值:

Vt=Rt+λAtplatform.V_t = R_t + \lambda A_t^{platform}.

若 policy:

ΔRt>0\Delta R_t>0

但:

ΔAtplatform<0,\Delta A_t^{platform}<0,

則短期財報可能改善,

但:

Vt+TV_{t+T}

下降。


16. Delayed Ecosystem Cost

本文定義 Delayed Ecosystem Cost:

DEC(π)=τ=1Tγτ[Hbaseline(t+τ)Hπ(t+τ)]+.DEC(\pi) = \sum_{\tau=1}^{T} \gamma^\tau [ H_{baseline}(t+\tau) - H_{\pi}(t+\tau) ]_+.

其中:

[]+[\cdot]_+

取 positive loss。

典型 DEC 包括:

  • creator exit;
  • newcomer decline;
  • content diversity loss;
  • user habit weakening;
  • advertiser ROI decline;
  • trust erosion。

這些通常不會在:

Day1Day1

A/B test 中完整出現。


17. Leading, Lagging, and Hidden Metrics

17.1 Leading metrics

快速變化:

CTR,SessionLength,WatchEvents.CTR, SessionLength, WatchEvents.

17.2 Lagging metrics

較慢顯現:

CreatorExit,CohortRetention,AdvertiserRenewal,OrganicEntry.CreatorExit, CohortRetention, AdvertiserRenewal, OrganicEntry.

17.3 Hidden-state metrics

難以直接測量:

Trust,Agency,InformationUtility,OpportunityCost,CreatorDiscouragement.Trust, Agency, InformationUtility, OpportunityCost, CreatorDiscouragement.

平台如果只優化:

LeadingMetrics,LeadingMetrics,

容易忽略:

LaggingLoss.LaggingLoss.

18. Metric-Induced Behavior Shift

當 creator 知道:

MetricxMetric_x

決定 exposure,

creator 會適應:

StrategykOptimize(Metricx).Strategy_k \rightarrow Optimize(Metric_x).

本文定義:

MIBS=CreatorBehaviorMetricIncentive.MIBS = \frac{ \partial CreatorBehavior }{ \partial MetricIncentive }.

例如:

CTRCTR

高度重要時,

可能增加:

ThumbnailOptimization,TitleOptimization,TopicConvergence.ThumbnailOptimization, TitleOptimization, TopicConvergence.

這不必然是壞事。

但:

MIBSMIBS

過強時:

ContentContent

開始服務 metric,而不是 audience value。


19. Platform Goodhart Loop

平台 Goodhart loop:

MetricChosenMetricChosen \Downarrow RankingOptimizesMetricRankingOptimizesMetric \Downarrow UIAndCreatorsAdaptUIAndCreatorsAdapt \Downarrow MetricDistributionChangesMetricDistributionChanges \Downarrow MetricBecomesLessInformativeAboutGoal.MetricBecomesLessInformativeAboutGoal.

形式:

Optimize(M)ChangeDataGeneratingProcessCorr(M,G).\boxed{ Optimize(M) \rightarrow ChangeDataGeneratingProcess \rightarrow Corr(M,G)\downarrow. }

這是推薦平台最重要的 proxy risk。


20. Stakeholder Utility Divergence

本文定義:

SUD=Var(ΔUs:sS).SUD = Var( \Delta U_s : s\in\mathcal{S} ).

若某改版:

UplatformshortU_{platform}^{short}\uparrow

但:

Ucreator,U_{creator}\downarrow, Uuser,U_{user}\downarrow,

則:

SUD.SUD\uparrow.

SUD 本身不代表 policy 必然錯。

但高:

SUDSUD

表示平台必須做顯式 governance trade-off,而不能稱為「全面改善」。


21. Product Health Gap

定義:

PHG=NormalizedMetricPerformanceNormalizedEcosystemHealth.PHG = NormalizedMetricPerformance - NormalizedEcosystemHealth.

如果:

PHG0,PHG\gg0,

表示 dashboard 看起來很健康,

但 ecosystem metrics 沒有同步改善。

這就是:

DashboardRealityGap.\boxed{ DashboardRealityGap. }

22. Metric Portfolio

單一 metric 容易 Goodhart。

因此平台應使用:

M={M1,,Mn}.\mathcal{M} = \{ M_1,\ldots,M_n \}.

例如:

Mengagement,M_{engagement}, Msatisfaction,M_{satisfaction}, Magency,M_{agency}, Mdiversity,M_{diversity}, Mcreator,M_{creator}, Madvertiser,M_{advertiser}, Mlongterm.M_{longterm}.

重要的是:

MetricPortfolioMetricPortfolio

也不能只變成更多 KPI。

必須有:

ExplicitTradeoffPolicy.\boxed{ ExplicitTradeoffPolicy. }

23. Guardrail Metrics

一個 ranking experiment 可以最大化:

PrimaryMetric.PrimaryMetric.

但需要 guardrails:

G={G1,,Gm}.G= \{ G_1,\ldots,G_m \}.

例如:

CTRCTR\uparrow

只有在:

ExplicitPreferenceViolationθ1,ExplicitPreferenceViolation\leq\theta_1, CreatorExposureGiniθ2,CreatorExposureGini\leq\theta_2, LongTailObservabilityθ3,LongTailObservability\geq\theta_3, UserSatisfactionθ4UserSatisfaction\geq\theta_4

時才接受。

因此:

Uplift without guardrails is not sufficient evidence of improvement.\boxed{ Uplift \text{ without guardrails is not sufficient evidence of improvement.} }

24. Hard Constraints versus Soft Objectives

不是所有價值都適合加權平均。

例如:

J=CTRλPrivacyViolationJ = CTR - \lambda PrivacyViolation

可能暗示:

HighCTRHighCTR

可以補償 privacy violation。

但某些條件應是:

PrivacyViolation=0.PrivacyViolation=0.

同理:

ExplicitBlockViolation=0ExplicitBlockViolation=0

應更接近 hard constraint。

因此推薦治理應區分:

SoftObjectiveSoftObjective

與:

HardConstraint.HardConstraint.

25. Long-Horizon Platform Utility

本文提出:

LHPU=t=0Tγt[wuUuser,t+wcUcreator,t+waUadvertiser,t+wpUplatform,t].\boxed{ LHPU = \sum_{t=0}^{T} \gamma^t [ w_uU_{user,t} + w_cU_{creator,t} + w_aU_{advertiser,t} + w_pU_{platform,t} ]. }

其中:

0<γ1.0<\gamma\leq1.

若:

γ0,\gamma\rightarrow0,

系統近似短期 extraction。

若:

γ\gamma

較高,

future supply、trust 與 retention 會更重要。


26. Dynamic Platform State

平台 state:

St=(Userst,Creatorst,Contentt,Relationst,Advertiserst,Trustt).S_t = ( Users_t, Creators_t, Content_t, Relations_t, Advertisers_t, Trust_t ).

policy:

πt.\pi_t.

狀態轉移:

St+1=F(St,πt,Externalt).S_{t+1} = F( S_t,\pi_t,External_t ).

因此推薦系統不是 static ranking problem。

它是:

ControlProblemOverPlatformState.\boxed{ ControlProblemOverPlatformState. }

27. Myopic Optimization

若平台只解:

πt=argmaxπRt(π),\pi_t^* = \arg\max_\pi R_t(\pi),

可能得到:

MyopicPolicy.MyopicPolicy.

更合理:

π=argmaxπE[τ=tt+TγτtU(Sτ)].\pi^* = \arg\max_\pi E[ \sum_{\tau=t}^{t+T} \gamma^{\tau-t} U(S_\tau) ].

這也是 recommender systems 從 supervised ranking 走向 sequential decision-making 的重要理由之一。


28. User Time as a Scarce Resource

平台常將:

WatchTimeWatchTime

視為收益。

但對 user:

TimeTime

是成本。

因此同一分鐘:

+1+1

對 platform engagement,

可能是:

1-1

對 user opportunity budget。

真正 user net utility:

NUu=ValueConsumedTimeCostAttentionCostOpportunityCost.NU_u = ValueConsumed - TimeCost - AttentionCost - OpportunityCost.

如果:

WatchTimeWatchTime\uparrow

但:

NUu,NU_u\downarrow,

則 engagement uplift 不等於 user welfare uplift。


29. Information Platforms and Freshness

對新聞、AI、金融、天氣等高時效資訊:

Value(v,t)=V0eλΔt.Value(v,t) = V_0e^{-\lambda\Delta t}.

若 recommender 偏好:

PopularOldContentPopularOldContent

而壓制:

FreshNicheContent,FreshNicheContent,

可能:

WatchTimeWatchTime

保持高,

但:

InformationUtility.InformationUtility\downarrow.

因此 domain-specific utility 必須納入:

Freshness.Freshness.

30. Surface-Level Product Contracts

A01 已區分:

Home,Following,Explore,Popular,Search.Home, Following, Explore, Popular, Search.

A06 進一步認為:

每個 surface 是:

ProductContract.\boxed{ ProductContract. }

例如:

FollowingFollowing

承諾較高 explicit relation fidelity。

ExploreExplore

承諾較高 novelty。

PopularPopular

承諾 global popularity。

如果全部 surface 最後都被:

EngagementMetricEngagementMetric

統一支配,

則:

SurfaceDifferentiation0.SurfaceDifferentiation\rightarrow0.

這是一種 product-contract collapse。


31. Multi-Surface Objective Separation

令:

JHJ_H

為 Home objective,

JFJ_F

為 Following objective,

JEJ_E

為 Explore objective,

JPJ_P

為 Popular objective,

JSJ_S

為 Search objective。

要求:

JHJFJEJPJS.J_H \neq J_F \neq J_E \neq J_P \neq J_S.

不是數學上必須完全不同,

而是 product intent 不應被單一:

JengagementJ_{engagement}

完全吞併。


32. A/B Testing Failure

標準 A/B test:

TreatmentΔMetric.Treatment \rightarrow \Delta Metric.

如果 window:

TABT_{AB}

很短,

但 ecosystem effect delay:

TecoTAB,T_{eco}\gg T_{AB},

則:

A/BA/B

只能觀察:

ShortTermEffect.ShortTermEffect.

因此:

NoShortTermHarm⇏NoLongTermHarm.\boxed{ NoShortTermHarm \not\Rightarrow NoLongTermHarm. }

33. Delayed Holdout

平台可以保留:

LongTermHoldoutCohort.LongTermHoldoutCohort.

對:

30,60,90,18030, 60, 90, 180

天觀察:

Retention,Satisfaction,CreatorDiversity,CreatorExit,AdValue.Retention, Satisfaction, CreatorDiversity, CreatorExit, AdValue.

這提供:

DECDEC

的估計。


34. Cohort Analysis

aggregate metrics 容易掩蓋:

UserGroupDifferences.UserGroupDifferences.

例如:

MainstreamUsersMainstreamUsers\uparrow

但:

NicheUsers.NicheUsers\downarrow.

總體:

WatchTime.WatchTime\uparrow.

因此需要:

ΔUmainstream,\Delta U_{mainstream}, ΔUniche,\Delta U_{niche}, ΔUhighIntent,\Delta U_{highIntent}, ΔUnew.\Delta U_{new}.

creator 同理。


35. Survivorship Bias

若:

CreatorsExitCreatorsExit

之後不再出現在 active dataset,

平台可能觀察:

AverageCreatorPerformance.AverageCreatorPerformance\uparrow.

這甚至可能只是:

LowPerformanceCreatorsRemovedFromDenominator.LowPerformanceCreatorsRemovedFromDenominator.

因此:

AveragePerformance⇏CreatorEcosystemImproved.\boxed{ AveragePerformance\uparrow \not\Rightarrow CreatorEcosystemImproved. }

36. Metric Decomposition

任何 metric uplift 應拆解來源。

例如:

ΔWatchTime=ΔWTbettermatch+ΔWTautoplay+ΔWTcontentlength+ΔWTrepeat+ΔWTfriction.\Delta WatchTime = \Delta WT_{bettermatch} + \Delta WT_{autoplay} + \Delta WT_{contentlength} + \Delta WT_{repeat} + \Delta WT_{friction}.

如果只看 total:

ΔWatchTime>0,\Delta WatchTime>0,

會失去 causal interpretation。


37. Product Success Criterion

本文提出最低 product success 條件:

policy π1\pi_1 相對 π0\pi_0,若要稱為 product improvement,至少需要:

ΔUMS0,\Delta U^{MS}\geq0,

以及:

HardConstraintsSatisfied=1,HardConstraintsSatisfied=1,

並且:

DECθ.DEC\leq\theta.

不要求所有 stakeholder:

ΔUs>0.\Delta U_s>0.

但 trade-off 必須被明確知道。


38. Platform Health Dashboard

合理 dashboard 至少包含六層。

Layer 1 — Engagement

CTR,WatchTime,ReturnRate.CTR, WatchTime, ReturnRate.

Layer 2 — Satisfaction

Survey,NotInterested,Complaint,VoluntaryReturn.Survey, NotInterested, Complaint, VoluntaryReturn.

Layer 3 — Agency

ExplicitPreferenceViolation,SearchCost,FollowingVisibility.ExplicitPreferenceViolation, SearchCost, FollowingVisibility.

Layer 4 — Exposure

Gini,HHI,LongTailObservability,ColdStartObservability.Gini, HHI, LongTailObservability, ColdStartObservability.

Layer 5 — Creator Ecosystem

CSTR,CVR,ExitRate,TTSA,SupplyHHI.CSTR, CVR, ExitRate, TTSA, SupplyHHI.

Layer 6 — Commercial Quality

QualifiedAttention,PaidROI,AdvertiserRenewal,Incrementality.QualifiedAttention, PaidROI, AdvertiserRenewal, Incrementality.

39. Product Health Gap Monitoring

每期計算:

PHGt=Score(Mt)Score(Ht).PHG_t = Score(M_t) - Score(H_t).

若:

MtM_t\uparrow

但:

Ht,H_t\downarrow,

則:

PHGt.PHG_t\uparrow.

應觸發:

MetricDivergenceReview.\boxed{ MetricDivergenceReview. }

40. Metric Governance

metric governance 不應只由 model team 決定。

需要:

  • recommender engineers;
  • product;
  • creator ecosystem;
  • advertising;
  • trust and safety;
  • research;
  • user research。

因為:

MetricChoiceMetricChoice

本身就是:

PlatformPolicy.PlatformPolicy.

41. Recommendation Governance as Constitutional Layer

可將推薦架構分為:

Layer 1 — Prediction

預測:

P(click),P(watch),P(save).P(click), P(watch), P(save).

Layer 2 — Ranking

生成:

Score.Score.

Layer 3 — Observation Policy

決定:

Home,Explore,FollowHome, Explore, Follow

如何分配 visibility。

Layer 4 — Governance

決定:

WhichMetricsMatter,WhichConstraintsCannotBeViolated,WhichStakeholdersCount.WhichMetricsMatter, WhichConstraintsCannotBeViolated, WhichStakeholdersCount.

A06 的核心是:

RankingCannotDefineItsOwnSuccessCriteria.\boxed{ RankingCannotDefineItsOwnSuccessCriteria. }

42. A Unified Series-A Dynamic Model

綜合 A01–A06:

CtObservationPolicyExposuret\mathcal{C}_t \xrightarrow{ ObservationPolicy } Exposure_t \Downarrow TelemetrytPreferenceInferenceUserModelt+1Telemetry_t \xrightarrow{ PreferenceInference } UserModel_{t+1} \Downarrow ExposureDistributiont+1ExposureDistribution_{t+1} \Downarrow CreatorReturnt+1CreatorReturn_{t+1} \Downarrow CreatorSupplyt+2CreatorSupply_{t+2} \Downarrow PlatformMetricst+2.PlatformMetrics_{t+2}.

然後組織使用:

PlatformMetricsPlatformMetrics

決定下一輪:

ObservationPolicy.ObservationPolicy.

因此形成:

PolicyObservationBehaviorSupplyMetricsPolicy.\boxed{ Policy \rightarrow Observation \rightarrow Behavior \rightarrow Supply \rightarrow Metrics \rightarrow Policy. }

這是一個完整 platform control loop。


43. The Metric-Success Trap

Metric-Success Trap:

  1. 更高 popularity prior;
  2. 高流量內容預測更穩定;
  3. CTR / watch time 上升;
  4. dashboard 判定成功;
  5. policy 繼續強化;
  6. long-tail / creator viability 下降;
  7. ecosystem cost 延遲出現;
  8. 到真正 business metric 下降時,供給結構已改變。

形式:

ShortTermMetricGainPolicyLockInDelayedStructuralLoss.\boxed{ ShortTermMetricGain \rightarrow PolicyLockIn \rightarrow DelayedStructuralLoss. }

44. Reversibility

平台 policy 應考慮:

Reversibility.Reversibility.

UI weighting 可以很快修改。

但 creator exit:

ExitkExit_k

可能不可逆。

community relation loss 也可能:

RecoveryCostPolicyChangeCost.RecoveryCost\gg PolicyChangeCost.

因此:

IrreversibleEcosystemEffects\boxed{ IrreversibleEcosystemEffects }

需要更高 experiment threshold。


45. Precaution for High-Leverage Metrics

若某 metric:

MM

同時影響:

  • ranking;
  • employee goals;
  • creator strategy;
  • ad pricing;
  • product experiments;

則其:

SystemLeverageSystemLeverage

非常高。

高 leverage metric 應要求:

HigherValidation.HigherValidation.

46. Practical Experiment Acceptance Rule

本文提出:

接受 treatment π1\pi_1 若:

PrimaryMetricUplift>0PrimaryMetricUplift>0

且:

GuardrailViolations=0GuardrailViolations=0

且:

ShortTermSatisfactionbaselineShortTermSatisfaction\geq baseline

且:

ExposureHealththreshold.ExposureHealth\geq threshold.

之後:

LongTermHoldoutLongTermHoldout

驗證:

DEC.DEC.

47. Case-Study Boundary

本 Series 最初由影音平台推薦體驗問題觸發,但本文不主張任何特定平台已發生全部 failure chain。

例如觀察到:

HighPopularityContentShareHighPopularityContentShare\uparrow

只足以建立:

Hypothesis.Hypothesis.

若要推出:

CreatorExitCreatorExit

或:

AdvertiserValueDecline,AdvertiserValueDecline,

需要後續 longitudinal evidence。

因此:

FrameworkVerdict.\boxed{ Framework \neq Verdict. }

Series A 的價值是建立:

WhatToMeasure,WhatToDistinguish,WhatCausalChainsToTest.WhatToMeasure, WhatToDistinguish, WhatCausalChainsToTest.

48. Design Recommendations

本文建議平台:

  1. 將 ranking metrics 與 platform health metrics 分層;
  2. 不把 watch history 視為同質 preference evidence;
  3. 保留 explicit preference semantics;
  4. 分離 Popular 與 Explore;
  5. 保留 long-tail / cold-start testability;
  6. 監測 creator cohort survival;
  7. 評估 paid promotion 是否轉化為 organic viability;
  8. 對 autoplay 與 passive exposure 做 provenance;
  9. 使用 multi-stakeholder utility;
  10. 建立 long-term holdout 與 delayed ecosystem audit;
  11. 對高不可逆性 policy 使用更高 deployment threshold;
  12. 允許 users 選擇 observation policy。

49. Limitations

第一,多 stakeholder utility 的權重:

wsw_s

具有治理與價值判斷,不存在唯一客觀最優解。

第二,部分 hidden-state metrics,例如 trust、agency 與 information utility,測量成本高且可能有 survey bias。

第三,長期 holdout 會降低短期 experiment velocity,平台需要在學習速度與長期識別能力之間取捨。

第四,Goodhart effects 並不表示所有 proxy 都無效。工程系統仍必須使用 metrics;問題是需要理解其邊界與資料生成機制。

第五,Series A 建立的是 general framework,不提供對特定平台演算法內部權重的證明。


50. Conclusion

Series A 從一個看似簡單的問題開始:

為什麼推薦頁一直給我不想看的東西?

最後得到的不是一個單純 accuracy 問題,而是一個完整的平台控制系統。

第一層:

Recommendation=ObservationPolicy.\boxed{ Recommendation = ObservationPolicy. }

第二層:

AwarenessInterestIntent.\boxed{ Awareness \neq Interest \neq Intent. }

第三層:

PlatformInducedBehaviorUserPreferenceEvidence.\boxed{ PlatformInducedBehavior \neq UserPreferenceEvidence. }

第四層:

TopicDiversityExposureDiversity.\boxed{ TopicDiversity \neq ExposureDiversity. }

第五層:

NoExposureLowQuality.\boxed{ NoExposure \neq LowQuality. }

第六層:

MetricSuccessProductSuccess.\boxed{ MetricSuccess \neq ProductSuccess. }

完整 Series A dynamics:

ObservationPreferenceInterpretationEndogenousContaminationExposureConcentrationCreatorSupplyResponseMetricGovernance.\boxed{ Observation \rightarrow PreferenceInterpretation \rightarrow EndogenousContamination \rightarrow ExposureConcentration \rightarrow CreatorSupplyResponse \rightarrow MetricGovernance. }

最終平台應優化的不是:

maxCTR\max CTR

也不是:

maxWatchTime.\max WatchTime.

而是:

maxLongTermMultiStakeholderUtility\boxed{ \max LongTermMultiStakeholderUtility }

subject to:

UserAgency,Safety,CreatorTestability,InformationQuality,CommercialIntegrity.\boxed{ UserAgency, Safety, CreatorTestability, InformationQuality, CommercialIntegrity. }

推薦系統不是一個單純替內容排序的函數。

它是一個:

DynamicVisibilityAllocationSystem.\boxed{ DynamicVisibilityAllocationSystem. }

而 visibility allocation 會改變:

Behavior,PreferenceData,Popularity,CreatorStrategy,ContentSupply,Revenue,PlatformFuture.Behavior, PreferenceData, Popularity, CreatorStrategy, ContentSupply, Revenue, PlatformFuture.

因此,推薦平台真正的工程挑戰不是:

如何讓模型更懂得讓使用者點擊?

而是:

HowCanAPlatformAllocateObservationWithoutDestroyingTheEcologyThatMakesObservationValuable?\boxed{ HowCanAPlatformAllocateObservationWithoutDestroyingTheEcologyThatMakesObservationValuable? }

這是 Series A 的最終命題。


References

[0] Neo.K / EveMissLab. “Recommendation as an Observation Operator.” Series A, Paper A01, v0.1, 2026.

[1] Neo.K / EveMissLab. “Recommendation, Cold Start, and Creator Ecological Collapse.” Series A, Paper A05, v0.1, 2026.

[2] McNee, S. M., Riedl, J., & Konstan, J. A. “Being Accurate is Not Enough: How Accuracy Metrics Have Hurt Recommender Systems.” CHI Extended Abstracts, 2006.

[3] Knijnenburg, B. P., Willemsen, M. C., Gantner, Z., Soncu, H., & Newell, C. “Explaining the User Experience of Recommender Systems.” User Modeling and User-Adapted Interaction, 22, 441–504, 2012. DOI: 10.1007/s11257-011-9118-4.

[4] Burke, R., Abdollahpouri, H., Mobasher, B., & Gupta, T. “Towards Multi-Stakeholder Utility Evaluation of Recommender Systems.” UMAP Extended Proceedings, 2016.

[5] Abdollahpouri, H., Burke, R., & Mobasher, B. “Recommender Systems as Multistakeholder Environments.” UMAP, 2017.

[6] 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.

[7] Wu, Y., Cao, J., Xu, G., & Tan, Y. “TFROM: A Two-sided Fairness-Aware Recommendation Model for Both Customers and Providers.” SIGIR 2021, pp. 1013–1022. DOI: 10.1145/3404835.3462882.

[8] Chaney, A. J. B., Stewart, B. M., & Engelhardt, B. E. “How Algorithmic Confounding in Recommendation Systems Increases Homogeneity and Decreases Utility.” RecSys 2018, pp. 224–232. DOI: 10.1145/3240323.3240370.

[9] Mansoury, M., Abdollahpouri, H., Pechenizkiy, M., Mobasher, B., & Burke, R. “Feedback Loop and Bias Amplification in Recommender Systems.” CIKM 2020, pp. 2145–2148. DOI: 10.1145/3340531.3412152.

[10] Manheim, D., & Garrabrant, S. “Categorizing Variants of Goodhart's Law.” arXiv:1803.04585, 2018.

[11] Jannach, D., & Jugovac, M. “Measuring the Business Value of Recommender Systems.” ACM Transactions on Management Information Systems, 10(4), 2019.

[12] Adomavicius, G., Bockstedt, J. C., Curley, S. P., & Zhang, J. “Recommender Systems, Ground Truth, and Preference Pollution.” AI Magazine, 43(2), 177–189, 2022. DOI: 10.1002/aaai.12055.

[13] 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.

[14] 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.


Series A Closure

Series A 正篇完成:

  • A01 — Recommendation as an Observation Operator
  • A02 — Explicit Preference versus Inferred Preference
  • A03 — Passive Exposure and Endogenous Preference Contamination
  • A04 — The Platform-Induced Exposure Bubble
  • A05 — Recommendation, Cold Start, and Creator Ecological Collapse
  • A06 — Metric Success, Product Failure

後續工程化文件:

  • TW-A — User-Controllable Recommendation & Observation Architecture