PaperlayerView exact revision on arXiv

Source-linked AI summary

OranSim: Simulating Social Media Marketing

Jianxiang Ma, Mingfu Zhang, Xiaocui Yang, Yichen Gao, Junzhao Huang, Yuesong Hou

arXiv:2609.28388v1cs.SI

TL;DR

Before launch, brands choose creative, creator, targeting, and budget while the consumers reached and their responses remain unobserved. OranSim simulates campaign-specific exposure and heterogeneous consumer responses, propagates responses through segments, and pairs scenarios using shared populations and aligned random numbers. Under the shared propagation parameters, M 14 rises from 1,834,405 to 3,599,734: reach grows to 2.00 times the baseline and response mass to 1.96 times.

  • Problem

    Before launch, brands choose creative, creator, targeting, and budget while the consumers reached and their responses remain unobserved.

    Source: §1 · p. 1

  • Method

    OranSim simulates campaign-specific exposure and heterogeneous consumer responses, propagates responses through segments, and pairs scenarios using shared populations and aligned random numbers.

    Source: §1 · p. 2, §3.3 · p. 4

  • Results

    Under the shared propagation parameters, M 14 rises from 1,834,405 to 3,599,734: reach grows to 2.00 times the baseline and response mass to 1.96 times.

    Source: §5.2 · p. 7

  • Takeaways & Limitations

    The representation and evaluation support prelaunch analysis of how campaigns reach consumers and how their responses accumulate over time.

    Source: §7 · p. 9

    Testing intervention differences and trajectory shapes on real platforms requires campaign-level observations.

    Source: §6 · p. 9

Abstract

from arXiv · show

Social simulation studies how individual behavior and social interaction produce collective outcomes. In social media marketing, campaign actions shape which consumers encounter the content and how they respond; these responses then spread through the population. We propose OranSim, a social simulation framework that connects creative, creator, targeting, and budget choices to this process. Heterogeneous consumers receive exposure according to content matching and platform allocation and generate initial responses, which propagate among 60 population segments. Candidate campaigns share the initial population and aligned random numbers, making their response trajectories comparable under action changes. In a controlled synthetic campaign, doubling the budget approximately doubles reach while lowering mean content match and engagement probability among the reached consumers; mean 14-day cumulative simulated response mass rises to 1.96 times the baseline. LightGBM predictors fitted to 39,000 historical RedNote notes estimate platform engagement with log-scale $R^2$ of 0.56--0.62 in five-fold cross-validation; a separate 12,154-note corpus supplies temporal, unseen-creator, and held-out-niche test splits. Public-data experiments evaluate policy value and audience ranking, and paired synthetic outcomes test counterfactual scoring. Together, scenario trajectories and engagement estimates support campaign selection according to a prespecified marketing objective. Code is available at https://github.com/OranAi-Ltd/oransim.

1 Introduction

Campaign choices shape both who sees marketing content and how those consumers respond, making prelaunch comparison difficult because the audience and its responses are not yet observed. OranSim connects these choices to exposure, response, and population propagation, then combines simulated trajectories with engagement estimates for campaign selection.

1 Introduction

  • Creative, creator, and targeting choices shape exposure ranking, while budget determines how many ranked consumers are reached and therefore changes audience composition.The resulting population response depends on both the number reached and their response tendencies.Source: §1 · p. 1
  • The framework addresses prelaunch comparison by adding creative, creator, targeting, and budget choices to a simulation of population dynamics.Aggregate market-response models and post-publication cascade models do not by themselves trace which consumers each prelaunch option reaches.Source: §1 · p. 2
  • OranSim maps marketing actions to content matching, exposure allocation, and reach, then propagates selected consumers’ initial responses between demographic segments.Candidate scenarios share their initial population, mechanism parameters, and individual random numbers.Source: §1 · p. 2
  • Campaign objectives use predicted reads, likes, collects, or comments alongside each scenario’s simulated trajectory.In the controlled case, predicted collects rank options first, with cumulative simulated response breaking ties.Source: §1 · p. 2
  • The paper evaluates engagement prediction on private RedNote notes, policy value and audience ranking on public data, and counterfactual scoring on paired synthetic outcomes.A controlled skincare campaign also examines how doubling budget changes reach, audience composition, and cumulative simulated response.Source: §1 · p. 2

2 Related Work

Prior work studies social behavior and interaction, consumer diffusion, and marketing outcomes and decisions through distinct modeling and evaluation approaches. OranSim applies a campaign-action perspective that connects an initial audience and its responses to subsequent population propagation.

Social simulation

  • Social simulators represent individual behavior and interaction, while world models support action-conditioned prediction at different scales.OranSim represents a campaign as an action whose consequences unfold through statistical individual responses and intergroup propagation.Source: §2 · p. 2

Consumer interaction and market diffusion

  • Agent-based marketing connects consumer heterogeneity and local interaction to aggregate diffusion, while seeding research links initial-recipient choice to participation and reach.OranSim uses campaign choices to determine the initial audience and its response, then computes propagation over that population.Source: §2 · p. 2

Outcome prediction and decision evaluation

  • Market-response models estimate aggregate outcomes, cascade models forecast from observed events, and off-policy and treatment-effect methods evaluate policies and audience rankings.OranSim’s evaluation uses outcome and decision measures.Source: §2 · p. 2

3 OranSim

OranSim combines a heterogeneous simulated society, exposure and response modeling, scenario rollouts, historical engagement prediction, and objective-specific prelaunch comparison. The framework represents individual responses and their propagation across demographic segments.

3.1 Simulated Society for Marketing

  • Consumers’ interests, demographics, activity, personality, and response states enter distinct matching, targeting, exposure, and response calculations.Age band, gender, and city tier define 60 segments, where demographic similarity informs intersegment influence.Source: §3.1 · p. 3

3.2 From Individual Exposure to Population Response

  • Exposure ranks consumers using matching and platform-related factors, and a larger budget extends selection down the ranking, changing reach and audience composition.Selected consumers’ logistic click and engagement propensities form paid inputs; a segmented Hawkes process combines those inputs with retained excitation and induced organic response.Source: §3.2 · pp. 3–4

3.3 Scenario Rollouts of Marketing Actions

  • Creative, creator, audience, budget, and platform-share choices enter scenario rollouts, which share the population and aligned random numbers to reduce population-resampling variation.Daily trajectories retain response timing, while M14 compares cumulative response over the same window.Source: §3.3 · p. 4

3.4 Observed Engagement Prediction

  • Four separate LightGBM regressors predict reads, likes, collects, and comments from historical notes, with separate outcomes allowing campaigns to prioritize an engagement objective.The feature map includes note text representations, creator and platform features, publication-time features, and note age.Source: §3.4 · p. 4

3.5 Prelaunch Comparison

  • Prelaunch records combine engagement predictions with rollout trajectories, and the campaign objective determines how candidates are compared.For a collect-first objective, candidates within the budget cap are ranked by predicted collects, then mean cumulative simulated response resolves ties.Source: §3.5 · pp. 4–5

4 Evaluation Protocol

The evaluation tests quantities used in campaign comparison through engagement prediction, controlled rollouts, public-log policy and audience evaluation, and paired-synthetic scoring.

4 Evaluation Protocol

  • Evaluation covers engagement prediction, controlled campaign rollouts, public-log content allocation and audience selection, and scoring under alternative synthetic scenarios.Source: §4 · p. 5

4.1 Prediction and Controlled Response

  • Prediction evaluation compares predictor configurations and structured temporal, creator, and niche holdouts; controlled rollouts record reach, content match, engagement probability, and paid and organic response.Source: §4.1 · p. 5

4.2 Off-Policy Value Estimation

  • Off-policy evaluation estimates policy value from logged outcomes and assignment probabilities using IPS, SNIPS, doubly robust, and Switch-DR estimators.Effective sample size describes weight concentration; Switch-DR applies residual correction below a weight threshold.Source: §4.2 · p. 5

4.3 Audience Ranking and Paired Outcomes

  • Audience ranking uses treatment-effect learners to order customers by estimated incremental response, with uplift@q, normalized AUUC, and Qini summarizing ranking performance.Source: §4.3 · p. 5
  • Paired synthetic tests assess factual fit, alternative-outcome error, effect direction, and grouped CATE fit under feature substitution.Source: §4.3 · p. 5

5 Experiments

Experiments test engagement prediction, controlled campaign rollouts, public-data decision evaluation, and paired counterfactual scoring. In the controlled campaign, doubling budget raises 14-day simulated response mass to 1.96 times baseline.

5.2 Simulated Social Response to Marketing Actions

  • Under the CNY 80,000 cap, the collect-first rule selects sk, which has the most predicted collects.Figure 2 shows four note predictions and daily simulated responses; budget changes leave predictions unchanged.Source: Figure 2 · p. 7, §5.2 · p. 7
  • Response mass rises to 1.96 times baseline when reach doubles to 2.00 times; the added audience has lower mean response propensity.The comparison uses shared propagation parameters; mean content match falls from 0.5009 to 0.4942, and mean engagement probability from 0.7020 to 0.6783.Source: §5.2 · p. 7, Table 3 · p. 7
  • At CNY 40,000, creative B preserves reach but lowers response mass from 1,834,405 to 1,832,978; the high-tier creator raises all four predicted counts.B receives 1,020 predicted collects versus A’s 1,371, while the high-tier creator yields 1,480 predicted collects and 79 comments.Source: §5.2 · p. 7

5.3 Evaluating Campaign Decisions on Public Data

  • On Open Bandit, estimates span 0.00398–0.00407 under random assignment and 0.00361–0.00432 under BTS; KuaiRand’s effective sample size is about 1/487 of its record count.KuaiRand’s SNIPS estimate is 0.0733, with interval [0.0618, 0.0850] and standard error 0.0055.Source: Table 4 · p. 8, §5.3 · p. 8
  • On X5, two-model logistic regression has the highest mean AUUC, X-Learner the highest Qini, and the preferred ranking method varies with campaign reach.DRLearner leads at 5% and 10% reach; X-Learner leads at 20% and 30%.Source: Table 4 · p. 8, §5.3 · p. 8

5.4 Paired Counterfactual Scoring

  • Feature-substitution scoring fits factual outcomes, then scores factual and alternative features with fixed predictors to estimate intervention effects.The 2,000 paired test scenarios evaluate factual outcomes, alternative outcomes, individual-effect direction, and grouped effects; grouped CATE fit averages generated and predicted differences within budget, creator, and niche groups.Source: Table 5 · p. 8, §5.4 · p. 8

6 Limitations

The engagement predictors model a private, observational RedNote corpus selected by engagement thresholds and within-niche rankings, while simulation response and propagation use design parameters. Testing intervention differences and trajectory shapes on real platforms requires campaign-level observations, and full predictor retraining depends on private data and an embedding service.

6 Limitations

  • The predictors represent a private observational RedNote corpus selected by engagement thresholds and within-niche rankings, limiting their direct scope to that corpus’s platform distribution.Raw notes, account identifiers, user profiles, and predictor weights remain private.Source: §6 · p. 8
  • Real-platform tests of intervention differences and response trajectories require campaign-level observations; individual conditional treatment effects remain unobservable in the X5 ranking data.The four note-outcome predictors are fitted to observations, whereas response, propagation, and intergroup influence use design parameters.Source: §6 · p. 9
  • Full retraining of the platform predictors requires the private corpus and original text-embedding service.The deterministic snapshot instead uses hashed text representations for scoring checks.Source: §6 · p. 9

7 Conclusion

OranSim connects marketing actions to population response through exposure allocation, heterogeneous individual behavior, and intergroup propagation, using shared initial populations and aligned random numbers to compare campaign scenarios. The controlled campaign, historical-note predictors, and public-data and synthetic experiments support prelaunch analysis of campaign reach and the accumulation of consumer responses over time.

7 Conclusion

  • OranSim connects marketing actions to population response through exposure allocation, heterogeneous individual behavior, and intergroup propagation.Shared initial populations and aligned random numbers support comparison of candidate campaign trajectories under action changes.Source: §7 · p. 9
  • Historical-note predictors provide engagement estimates that campaign objectives use alongside simulated rollout trajectories for scenario-specific selection.The prediction, public-data, and paired synthetic experiments evaluate quantities used in these comparisons.Source: §7 · p. 9
  • Together, the representation and evaluation support prelaunch analysis of how campaigns reach consumers and how responses accumulate over time.Source: §7 · p. 9

A Simulator Specification

The simulator represents heterogeneous consumers, ranks them for exposure, models individual responses, and propagates those responses across demographic segments. In a controlled skincare campaign, doubling budget expanded reach and increased cumulative simulated response mass; a collect-first rule selected the option with the highest predicted collects.

A.1 Simulated Population and Segments

  • The population comprises 100,000 sampled individuals representing 2,000,000 people, organized into 60 age-gender-city segments for population-level propagation.Exposure and individual response are computed for individuals; propagation is computed between segments.Source: §A.1 · p. 12

A.2 Exposure Scoring and Budget Conversion

  • Exposure ranks consumers using content, activity, targeting, creator, platform-audience, and seeded-exploration factors; budget sets how many are selected.For the CNY 40,000 case, the model selected 41,666 individuals, representing a reach of 833,320 people.Source: §A.2 · p. 12

A.3 Individual Response and Paid Schedule

  • For each selected individual, initial response mass entering propagation is the click probability, and individual masses are summed within segments before distribution over 14 days by the paid schedule.Day 1 receives 33.1% of paid response mass, and the first two days receive 55.3%.Source: §A.3 · p. 13

A.4 Segment-Level Propagation

  • Intergroup influence is strongest within a demographic segment and weakens as age, gender, and city-tier differences increase.Source: §A.4 · p. 13

A.5 Rollout Procedure

  • Each rollout uses shared population and mechanism settings, selects exposed consumers, computes responses, injects paid mass, propagates organic response, and aggregates 14-day trajectories.Candidate scenarios differ in action settings while the population and per-seed latent perturbations stay fixed.Source: §A.5 · p. 13

B.2 Reach, Response Tendencies, and Cumulative Response

  • Under creative B, raising budget from CNY 40,000 to CNY 80,000 increased cumulative simulated response mass from 1,832,978 to 3,604,845 while mean engagement probability fell from 0.7009 to 0.6799.The number selected increased from 41,666 to 83,333; mean content match changed from 0.4971 to 0.4987.Source: §B.2 · p. 14

B.3 Daily Paid and Organic Response

  • Across options, paid response peaks on day 1 while organic response peaks on day 3; the CNY 80,000 options’ daily responses are about 1.96 times those of the CNY 40,000 options.The campaign options vary specified creative or budget fields against the baseline.Source: §B.3 · p. 14, §B.1 · p. 14

B.4 Applying the Collect-First Rule

  • The collect-first rule ranked options by predicted collects, used cumulative simulated response to break ties, and selected sk.The complete order was sk, sb, s0, scb, sc; predictions were identical for options sharing creative and creator but differing in budget.Source: §B.4 · p. 15

B.5 Seed Variation and Reproducibility Checks

The appendix reports seed variation and reproducibility checks, then describes prediction corpora and structured holdout splits, as well as policy-value estimators and their identification conditions. The supplied passages report seed-level variation but do not provide the prediction or policy-evaluation result values.

B.5 Seed Variation and Reproducibility Checks

  • Across 30 shared seeds, M14 standard deviations ranged from 992 to 1,299; mean difference from the creative change at CNY 40,000 was −1,427, while budget doubling yielded 1,765,329.Repeating the baseline with the same seed gave a maximum absolute difference of zero in aggregate and daily outputs; an empty intervention gave a zero difference.Source: §B.5 · p. 15

C.1 Corpora and Features

  • The shared-fold comparison uses 39,000 note snapshots, while a separate 12,154-note corpus supports structured holdout experiments.The comparison evaluates configurations without and with note age on shared shuffled folds.Source: §C.1 · p. 17

C.2 Holdout Splits

  • Structured holdouts test later-published notes, 1,642 creators absent from training, and pooled predictions from leave-one-niche-out folds across 13 niches.Preprocessing and predictors are refitted within each training split.Source: §C.2 · p. 17, §C.1 · p. 17, Table C.1 · p. 18

D.1 Off-Policy Estimators and Identification

  • DR and Switch-DR combine importance weighting with an outcome model; Switch-DR applies residual correction only when the weight is at most τ.The passage states that policy-value identification requires overlap, unconfoundedness, and the stable unit treatment value assumption.Source: §D.1 · p. 18

D.2 Uplift Metrics

The appendices define uplift-ranking metrics, compare audience-ranking methods at different reach fractions on X5 RetailHero, and describe the synthetic scenarios and splits used for counterfactual evaluation.

D.2 Uplift Metrics

  • Test customers are sorted by predicted effect in decreasing order before the gain curve evaluates audience ranking.Figure D.1 plots estimated incremental responses per 10,000 customers reached against the fraction of top-ranked customers reached; lines average three seeds and bands show ±1 standard deviation.Source: §D.2 · p. 18, Figure D.1 · p. 19

D.3 Case Study: Audience Ranking under Limited Reach

  • DRLearner ranks highest at 5% and 10% reach, while X-Learner leads from 20% to 50%; all methods approach random-ranking performance as reach expands.Figure D.1 reports estimated incremental responses per 10,000 customers reached; Table D.1 reports means and standard deviations over three seeds.Source: Figure D.1 · p. 19, §D.3 · p. 19, Table D.1 · p. 20

D.2 Uplift Metrics

  • Normalized AUUC compares gain-curve area with random and best-ordering areas, assigning 0 to random ordering and 1 to the best ordering.Qini replaces the gain curve's U(m) with its own measure and is normalized the same way; uplift-at-reach reports uplift at a specified reach fraction.Source: §D.2 · p. 19

E.1 Scenario Generator and Splits

  • The synthetic generator creates 20,000 scenarios; an alternative shifts the treatment arm and scales budget by the destination arm's factor, while identity interventions have exactly zero effect.Scenarios are split by identifier into 16,000 training, 2,000 validation, and 2,000 test cases, without overlap.Source: §E.1 · p. 20
  • The feature-substitution baseline fits a separate LightGBM regressor for each outcome using log(1 + y) as the target.Inputs include platform, niche, budget and creator tiers, treatment arm, log budget, log follower count, and creator engagement rate.Source: §E.1 · p. 20

E.2 Full Metrics and Segment-Level Effects

The section reports the metrics used to assess the LightGBM feature-substitution baseline and describes grouped-effect evaluation across budget tiers, creator tiers, and niches.

E.2 Full Metrics and Segment-Level Effects

  • Grouped CATE R2 is computed across 18 groups: four budget tiers, four creator tiers, and ten niches.Table E.1 extends Table 5 with these segment-level evaluations.Source: §E.2 · p. 20, Table E.1 · p. 21
  • Table E.1 reports factual-outcome, counterfactual-outcome, and individual-effect mean absolute errors, plus individual-effect rank correlation on 2,000 paired synthetic test scenarios.Counterfactual normalized MAE divides counterfactual MAE by the mean absolute alternative outcome.Source: §E.2 · p. 20, Table E.1 · p. 21
  • Figure E.1 compares generated and predicted group effects, plotting within-group mean generated effects against within-group mean predicted effects with equality marked diagonally.The plotted groups are the four budget tiers, four creator tiers, and ten niches.Source: §E.2 · p. 20, Figure E.1 · p. 21

F Reproducibility Materials

The reproducibility materials describe a deterministic engagement-prediction snapshot and identify which code, scenarios, experiment outputs, and source materials are publicly available.

F.1 Deterministic Output Snapshot

  • The output snapshot contains 3,000 scenario scoring records across three seeds and records four predicted engagement counts with per-follower rates.Records span 15 niches, three creator tiers, three budget tiers, two exposure regimes, and image and video posts; identical seeds and record counts regenerate byte-identical files.Source: §F.1 · p. 22, Table F.1 · p. 22
  • The snapshot's exposure field is set to eight times predicted reads, but the paper does not use that field.The snapshot also stores predictor version and seed for each record.Source: §F.1 · p. 22

F.2 Released and Private Materials

  • The public repository includes simulator code, synthetic scenarios, experiment scripts, and aggregate results for KuaiRand-Pure, Open Bandit, X5, and paired counterfactual experiments.Raw RedNote notes, account identifiers, and user profiles are not released.Source: §F.2 · p. 22
Loading 2609.28388v1…