acceptodds
Under review as a conference paper at ICLR 2027

OranSim: Simulating Consumer Response to Social Media Campaigns Before Launch

Abstract

A social media campaign must be chosen before its population response is observed. Creative, creator, targeting, and budget choices shape who encounters the content and how they respond; propagation turns these initial responses into population dynamics. We propose OranSim, a social simulation framework for social media marketing. It maps marketing actions to content matching, exposure allocation, and reach, computes individual responses, and propagates them among 60 population segments. Candidate scenarios share the initial population and random numbers to compare responses to action changes. In a controlled synthetic case, doubling the budget roughly doubles reach, lowers mean content match and engagement probability, and raises the mean 14-day cumulative simulated response mass to 1.96 times the baseline. For platform engagement objectives, predictors trained on historical notes estimate read, like, collect (save), and comment counts. On 39,000 private RedNote notes, they achieve log-scale of 0.56–0.62 in five-fold cross-validation; a separate set of 12,154 notes tests temporal, unseen-creator, and leave-one-niche-out prediction. Under a collect-first rule, predicted collects select the creative and simulated response selects the budget. Experiments on KuaiRand-Pure, X5 RetailHero, and the Open Bandit Dataset evaluate policy value and audience ranking under their assignment mechanisms, and paired synthetic outcomes test counterfactual scoring. By tracing how actions shape exposure, individual response, and propagation, OranSim supports prelaunch comparison and selection of campaigns for specific objectives.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
social simulation, social media marketing, agent-based modeling, Hawkes process, off-policy evaluation, uplift modeling

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_RXJhOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2(+1)
Read review
Paper summary

OranSim is a pre-launch simulator for social-media marketing campaigns on RedNote. A campaign (creative, creator, targeting, budget) selects consumers from a synthetic population of 100,000 using a multiplicative exposure score. A hand-specified logistic model gives each selected consumer's click and engagement probability, and a 60-segment Hawkes process propagates the response. Separately, LightGBM regressors trained on 39,000 private RedNote notes predict reads, likes, collects and comments, and a lexicographic rule combines the two to choose a campaign. The evaluation covers the LightGBM predictors, one synthetic skincare campaign, off-policy and uplift estimators on public datasets, and a synthetic counterfactual test.

Critical strengths
  1. The framework is interpretable. It links each marketing action to consumer exposure, individual response and population propagation. Because candidate campaigns share the population and the random numbers, they can be compared directly and the comparison can be reproduced.
Critical weaknesses
  1. The simulator is never validated against real consumer or campaign data, and its main result follows from its design. The population, exposure factors, logistic coefficients and Hawkes parameters are all set by hand (Table A.1), and the authors acknowledge that calibrating them would need campaign-level observations they do not have (Sec. 6). With these choices the headline finding (doubling the budget gives 1.96 times the response) is close to arithmetic: reach doubles by Eq. 1, and App. A.4 shows that total organic response is a fixed multiple of the paid input, whatever the influence matrix A. Replacing individual features by the audience mean changes M14 by less than 0.2% (Table 3), so consumer heterogeneity (Big Five, response state) adds little beyond who gets selected. Content vectors are hashed titles with no semantic relation to interests, so creatives A and B differ in M14 by under 0.1%, and the choice of creative is decided by LightGBM alone.

  2. The experiments do not test OranSim as a system, and it is unclear what the population rollout contributes to the decision. Sections 5.3 and 5.4 evaluate standard off-policy estimators and uplift learners on KuaiRand, Open Bandit and X5, none of which feed the rollout, plus a synthetic generator the authors built themselves. There is no baseline for the full framework, such as LightGBM or a market-response model on its own, or an existing social simulator. Under the collect-first objective the final choice rests on LightGBM, and the rollout only orders options that differ in budget alone. The predictors also transfer poorly over time (latest-time Rlog⁡2R^2_{\log} of 0.25 to 0.36, and −0.01 for comments) and under-predict counts by 57% to 82% in the structured holdouts.

Questions for authors
  1. Can you compare the rollout's predicted response with real campaign outcomes (for example paid RedNote campaigns with known budgets), and report the error?
  2. What does the rollout add to the decision beyond LightGBM? An ablation that selects campaigns with LightGBM alone, or with a simple reach × mean-propensity model, would show this. Does LightGBM beat simpler predictors?
  3. Predicted engagement does not depend on budget by construction (ψ(s)\psi(s) excludes it), so two campaigns with the same content get the same predicted collects at any budget. How should an advertiser use these predictions when deciding on a budget?
  4. Where do the demographic marginals (A.1) and the RedNote audience priors (A.2) come from, and how does the synthetic population compare with RedNote's actual user distribution?
  5. The hashed content vectors have no semantic relation to the interest vectors (Sec. 6). What, then, does content matching measure in the rollout?
  6. What is the basis for the logistic coefficients, the 0.5 engagement weight, the targeting strength of 2.0, the platform priors, and β=0.9\beta = 0.9, r=0.35r = 0.35? App. B.7 varies them one at a time; were any of them fitted to data?
  7. Please add fold-level variance for Table 1.
5 comments
AuthorsOpenReview ↗

Response to Reviewer RXJh (1/3)

Thank you for working through Eq. 1 and App. A.4 in this detail. W1 and W2 below refer to the two paragraphs of your weaknesses.

Q1 and W1: comparison with campaign outcomes. We ran this comparison on a paid TikTok campaign with known daily spend per creative (general response, item 1). With response intercepts and spend-to-reach conversion re-estimated on two weeks, the rollout predicts 1.08 times the observed engaged views of the held-out week (95% interval 0.98–1.21), with a creative-level absolute error of 0.26 of the observed total; its ratios for likes and 2-second views are 0.92 and 0.94. Reach × mean propensity gives 2.05 and 1.05, and LightGBM alone 1.26 and 0.34. The slopes, β and r in this test are the paper's design values. The components were also tested one at a time:

ComponentObserved dataResult
Population marginals (A.1)follower profiles of 12,983 RedNote creatorsgender: total-variation distance 0.005; age: 0.235, simulated population older
Response model (A.3)KuaiRand randomized exposure, held-out usersintercept re-estimation matches mean rates (ratios 1.005 click, 0.997 engagement); refitted AUC up to 0.688 / 0.807
Propagation timing (Eq. 2)489 RedNote notes observed from day 0–1 to day 3 or latercumulative-share RMSE 0.034 / 0.031 / 0.048 / 0.105 (reads / likes / collects / comments)
Predictor g_θ (§3.4)39k notes, note-disjoint temporal splitR² 0.443–0.575, Spearman ρ 0.671–0.714

The revision adds the campaign test as §5.5 and this table to App. A, and rewrites §6 to state which parts are re-estimated on the campaign, which are compared with observed data, and which remain design values.

Q6. Basis of the parameters. The logistic coefficients, the 0.5 engagement weight, the 2.0 targeting strength, the platform priors, β and r are design values; App. B.7 varies them one at a time, and the paper fits none of them. The new experiments fit the two parts that public or internal data can identify.

  • Response. KuaiRand-Pure is the public log with randomized exposure and per-exposure responses, so we fitted A.3's logistic structure on it with held-out users. Content match is proxied by tag cosine and fatigue by exposures in the previous 24 hours; targeting, personality and celebrity have no KuaiRand counterpart and stay constant. Results:
ModelClick AUCEngagement AUCPredicted / observed mean (click; engagement)
Design slopes, re-estimated intercepts0.5070.5581.005; 0.997
Same proxies, refitted0.5410.6701.006; 0.985
Plus Kuaishou viewing history and video attributes0.6880.8071.004; 1.083

The intercepts, which set absolute response levels, are identified from randomized exposure, and the same logistic structure takes the viewing history and video attributes without change. The revision reports this fit in App. A.3 and labels the slopes in Table A.1 as design values.

  • Timing. Fitting a single exponential accumulation rate to the 489 notes gives 0.11 (reads), 0.21 (likes), 0.27 (collects) and 0.46 (comments) per day; the rollout's paid schedule releases 0.4 per day, and propagation adds the slower tail. With propagation, the rollout's cumulative curve is closer to the observed median than the paid schedule alone for reads (RMSE 0.034 against 0.100) and likes (0.031 against 0.069), equal for collects (0.048 against 0.050), and comments follow the faster paid schedule (0.047 against 0.105). The revision reports these rates beside β and r in App. A.4.
AuthorsOpenReview ↗

Response to Reviewer RXJh (2/3)

W1: the 1.96 ratio. Your derivation is right, and it holds exactly. Total 14-day response is M₁₄ = w·K·m̄·F: K is reach, m̄ the mean initial response of the consumers that exposure ranking selects, w = 20 the population weight, and F = 1.7575 the propagation multiplier, which depends only on the paid schedule, β, r and the time step because A is row-normalized. The identity reproduces all 450 rollouts (3 populations × 30 seeds × 5 options) with relative error below 6×10⁻⁸. So 1.96 = 2.00 × 0.981: reach doubles, and the 41,667 added consumers, ranked lower by exposure score, respond less on average. §5.2 traces this composition effect (contribution 2: doubling the budget "expands reach, changes audience composition"). The revision writes the identity in App. A.4 and gives the 0.981 factor in §5.2.

W1: heterogeneity. §5.2 reaches the same reading: "Heterogeneity thus shapes aggregate response mainly through which consumers the exposure ranking selects." Table 3 isolates this channel, +4.4% to +6.6% over uniform selection and under 0.2% from substituting mean features, and the closed form above makes it exact for total response.

Q2 and W2: what the rollout adds to the decision. The rollout handles the two actions for which the predictor has no input, budget and targeting: LightGBM gives s_b the same prediction as s₀, and s_cb the same as s_c. We ran the two ablations you propose, on 3 populations × 30 seeds; the comparison of LightGBM with simpler predictors is in (3/3).

ModelError in total M₁₄Ranking by M₁₄Ranking by M₁₄ per CNY
Reach scaling from s₀0 to +1.93%options with the same budget tieall options tie
Reach × population-average response−4.2% to −6.2%same top option, order differsdifferent top option; budget pairs tie
Exposure ranking + initial responses, closed-form F< 6×10⁻⁸identicalidentical

The part of the rollout that changes decisions is the exposure ranking together with the individual responses of the consumers it selects. Removing it, as the population-average model does, makes response linear in reach, so under the response-per-CNY objective every budget pair ties and the model cannot choose a budget. Propagation sets the multiplier F and distributes response across segments and days: the cumulative segment shares of response differ from the selected audience's composition by a total-variation distance of 0.05–0.10, and replacing A with uniform weights moves the targeted segments' organic share from 96.1% to 94.3%.

Under the collect-first objective, ordering by predicted collects and then by budget gives the same ranking as Eq. 7 (s_k > s_b > s₀ > s_cb > s_c). As §5.2 states, the predictor chooses creative and creator there, and the rollout orders the budget levels. Under response objectives the rollout chooses: s_cb by M₁₄ and s₀ by M₁₄ per CNY.

On the paid campaign, the two models you name are the baselines in the general response (item 1). For the held-out week, LightGBM alone predicts 1.26 times the observed engaged views and reach × mean propensity 2.05 times, against 1.08 for the rollout, which takes each ad group's targeting and spend. The rollout's creative-level error is below LightGBM's by 0.08 of the observed total (95% interval 0.02–0.15).

Q3: budget and the predictor. ψ excludes b because the historical notes do not record promotion spend: the predictor estimates engagement from content, creator and timing, and the budget enters through reach in the rollout. An advertiser uses the predictor to choose content and creator and the rollout to choose the budget level. In the case, doubling the budget raises M₁₄ by 96.2% and lowers response per CNY by 1.9%; the size of this curvature is set by the response slopes, which are design values (Q6). The revision writes this division into §3.5, next to Eq. 7.

AuthorsOpenReview ↗

Response to Reviewer RXJh (3/3)

Q5 and W1: content matching. As §6 states, the rollout's content vectors are hashed titles, so content match in the rollout has no semantic relation to interests, and creatives A and B differ by less than 0.1% in M₁₄. In the case, creative choice is made by the predictor, which embeds title and body with text-embedding-3-small and whose temporal and grouped results are under Q7 below. The content-matching term in Eq. 1 is where an interest model with semantic vectors enters exposure ranking. The revision states in §3.2 and §5.2 that creative differences in the case come from the predictor, and that the rollout's content channel uses hashed vectors.

Q4: population sources. Age band, gender and city-tier marginals approximate China's Seventh National Census (2020). Platform penetration and audience priors approximate QuestMobile's 2024 reports, for example 65% female among RedNote users against the census share of 48.7%. App. A.1–A.2 will list every marginal and prior with its source. Against the follower profiles of 12,983 RedNote creators, the simulated RedNote users are 81.0–81.3% female against 80.8–83.4% (total-variation distance 0.005); QuestMobile's 65% counts all users, while the profiles count users who follow creators. The simulated age distribution is older: under 25 / 25–34 / 35–44 / 45+ is 21.0 / 26.7 / 30.0 / 22.4% against 37.4 / 33.8 / 19.4 / 9.5% (distance 0.235). The revision reports the age deviation in §6 and gives the profile marginals in App. A.1.

W2: experiments and framework baselines. §5.3 and §5.4 evaluate the estimators a campaign comparison uses when logs exist: off-policy estimators for content allocation, treatment-effect learners for audience selection, and predictor-based counterfactual scoring. They do not run the rollout. The revision adds the experiment–component table from the general response to §4. Framework-level baselines are the predictor alone and reach × population-average response, both under Q2. The paid campaign has spend and outcome series, so a market-response model joins these two baselines there (general response, item 1). As §1 notes, the social simulators we cite do not take creative, creator, targeting and budget as actions, so a comparison with them would first require adding this action space; the two decompositions in Q2 isolate what the rollout contributes.

Q7, the LightGBM part of Q2, and transfer over time (W2). Fold standard deviations: with Table 1's row split, 0.007–0.012; with folds grouped by note, R² is 0.504 / 0.539 / 0.482 / 0.419 (reads / likes / collects / comments) with standard deviations of 0.013–0.018. The revision uses the grouped protocol in Table 1 and the abstract. Simpler predictors on the same grouped folds:

PredictorReadsLikesCollectsComments
Training mean−0.000−0.000−0.000−0.000
log follower count, linear0.0050.0610.0540.021
Metadata only, LightGBM0.4290.4520.3890.276
Text embeddings only, LightGBM0.2600.2810.2660.284
Full predictor0.5040.5390.4820.419

On the note-disjoint temporal split of the 39k corpus, the full predictor reaches R² 0.562 / 0.575 / 0.503 / 0.443 and Spearman ρ 0.671–0.714, and it is higher than every baseline above and the creator's historical median on all four counts. The temporal row of Table 1 uses the separate 12k corpus with hashed text; there, comment R² of −0.007 is measured against the test-period mean, which is unavailable at deployment, and the training-period baselines score −1.01 to −5.23 on the same split. Back-transformed counts remain below observed means (44–59% of observed on the 39k temporal split); the collect-first rule uses only their order, as §6 states.

Reviewer_2RS4OpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The study introduces OranSim, integrating heterogeneous consumer behavior, campaign design, audience targeting, budget allocation, and intergroup information propagation to simulate consumer responses and compare marketing strategies before launch.

The study combines engagement prediction using 39,000 RedNote posts with controlled simulations and public-data evaluations, demonstrating how campaign choices influence reach, engagement, and diffusion, thereby supporting objective-driven marketing decisions.

Critical strengths

The paper combines heterogeneous consumer behavior, exposure allocation, and social diffusion with real-world engagement prediction, enabling systematic prelaunch comparisons of marketing strategies under shared initial conditions. This strengthens the interpretability and reproducibility of campaign evaluations.

Critical weaknesses

The engagement prediction models exhibit substantially degraded performance on temporally held-out data, including negative R^2 for comment prediction. This raises concerns about the framework’s robustness to evolving consumer preferences, platform dynamics, and changing marketing trends.

Questions for authors

What can you do to improve the performance on temporally held-out data?

1 comment
AuthorsOpenReview ↗

Response to Reviewer 2RS4

Thank you for the review. The temporal degradation in Table 1 is measured on the separate 12k corpus with hashed text features. With the semantic text features of the predictor used in the case, a temporal split over a one-week test window shows no degradation.

W1. Temporal robustness. We split the 39k RedNote corpus by publication date: training notes published 2026-03-17 to 04-13 (20,890 notes), test notes 04-14 to 04-20 (3,652 notes, none in training). Preprocessing is fitted on the training notes only.

PredictorReadsLikesCollectsCommentsCollects, Spearman ρ
Training mean−0.046−0.041−0.052−0.031—
log follower count, linear−0.0390.0120.000−0.0180.245
Creator's historical median0.3030.3280.2740.2040.534
Metadata only0.5010.5000.4220.3250.589
Text embeddings only0.2590.2470.2360.2770.502
Full predictor0.5620.5750.5030.4430.671

Values are log-scale R² except the last column. Comment R² is 0.443, and within-niche Spearman ρ is 0.514–0.569 across the four counts. On note-grouped cross-validation the same predictor scores 0.419–0.539, so over this horizon performance on later notes matches performance on unseen notes. The 12k corpus differs from the 39k corpus in sampling as well as in text features, so we do not attribute the whole difference between the two temporal results to the features.

On the 12k temporal split, the comment R² of −0.007 is measured against the test-period mean, which is unavailable at deployment. Training-period baselines score −1.01 (creator's historical median) to −5.23 (niche mean) on the same split, and the predictor's R² and MAE are better than every training-period baseline for all four counts.

Q1. Improving temporal performance. The main change is the semantic-feature predictor above, which shows no temporal degradation on its corpus. On the 39k temporal split we also tested three changes registered before running: recency weighting with 14- and 60-day half-lives changes R² by −0.003 to +0.007; a follower-normalized target lowers R² by 0.001–0.008; a Tweedie objective brings predicted mean counts closer to observed (from 44–59% to 68–82% of observed) at a cost in R². For use over time, retraining on a rolling window of recent notes keeps the training period adjacent to the notes being scored, and split-conformal intervals monitor drift: at the 90% level, log-scale intervals keep 85.7–86.8% coverage on the 12k temporal split, against about 90% on random splits. The revision adds the 39k temporal split and this coverage to Table 1, and the three changes to App. C.

Reviewer_4gDyOpenReview ↗
Rating
2
Confidence
1
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The authors propose a method of simulated consumer response to social media campaigns. It uses the idea of simulating a population of heterogeneous consumers, and then they add an exposure mechanism, individual responses, and hawkes process for simulation of response propagation across demographics. The study predicts engagement metrics such as reads, likes, collects, and comments, and it is trained on historical RedNote data. They test of skincare product and show that chanding the budget effect reach and their prediction model has moderate performance.

Critical strengths

the study can be practical and interesting, and the proposed method is structured and clean. paper is well organized and easy to follow iwth detailed information provided. also evaluations are on multiple dataset and with various aspects of prediction for decision making.

The paper addresses an interesting and practically relevant problem. Understanding how different campaign choices may affect consumer engagement before launching a campaign is useful, and the proposed framework provides a structured way to connect campaign decisions to exposure, individual responses, and aggregate outcomes.

Critical weaknesses

the motivation of the paper is that to simulate consumer responses to campaigns before launch and it would help select better campaign strategies. but, all experiments to prove this capability are entirely synthetic.

the confusing part is that paper includes several experiments using RedNote, KuaiRand, Open Bandit, X5 RetailHero, and synthetic counterfactual data but at the end of the day each experiment check a different prediction or estimation method rather than the integrated OranSim framework.

Questions for authors

i was wondring if authors can test on real-world campaign data? it can be small experiment, but would help see if claims are supported.

i could not find the model selection process for the Hawkes process and response model.

i could not follow how public-data experiments validate the OranSim framework. i mean specifically can u clarify the components that off-policy estimation and uplift-modeling experiments could evaluate?

1 comment
AuthorsOpenReview ↗

Response to Reviewer 4gDy

Thank you for the review. The general response gives the full numbers for the new experiments cited below.

W1. Synthetic and observed evidence. The campaign case (§5.2) and the paired scoring test (§5.4) are synthetic. The other experiments use observed data: §5.1 fits and tests the engagement predictor on 39k and 12k RedNote notes, and §5.3 evaluates estimators on KuaiRand, Open Bandit and X5 logs. The new experiments add a held-out test of the rollout on a paid campaign (Q1) and observed data for three rollout components:

  • population marginals against the follower profiles of 12,983 RedNote creators (gender total-variation distance 0.005);
  • the logistic response model on KuaiRand's randomized exposures, where re-estimated intercepts match mean click and engagement rates (ratios 1.005 and 0.997);
  • the rollout's early accumulation against 489 RedNote notes observed from their first day, within RMSE 0.034 / 0.031 / 0.048 of the median curve for reads / likes / collects.

The predictor itself is tested over time on a note-disjoint temporal split of the 39k corpus: R² 0.443–0.575, Spearman ρ 0.671–0.714, higher than follower-count, creator-history, metadata-only and text-only baselines on all four counts.

W2 / Q3. What each experiment tests. The table in the general response (item 6) maps each experiment to the component it tests. §5.3 answers the question a campaign comparison faces once exposure or treatment logs exist: which estimator gives the policy value of a content allocation (off-policy estimators on KuaiRand and Open Bandit), and which learner ranks customers for targeting (treatment-effect learners on X5). These estimators evaluate the allocation and targeting inputs; they do not run the rollout. The integrated pipeline runs in §5.2, where the frozen predictor and the rollout rank the five campaign options together, and in the new campaign test, where the rollout is compared with the predictor alone on held-out outcomes; the new ablations in our response to RXJh (Q2) show which part of the pipeline decides each choice. The revision adds the table to §4 and says this in §5.3.

Q1. A small real campaign. We ran this test on a paid TikTok campaign by a consumer-electronics brand: 691 ad creatives, USD 769 of spend and 1.48M impressions over three weeks. With its response intercepts and spend-to-reach conversion re-estimated on the first two weeks, the rollout predicts 1.08 times the observed engaged views of the held-out third week, with a creative-level absolute error of 0.26 of the observed total. A model that multiplies reach by the mean response predicts 2.05 times the observed total, and a LightGBM predictor alone 1.26 times. The general response (item 1) gives the table, and the revision adds the test as §5.5.

Q2. Model selection for the Hawkes process and the response model.

  • Response model: on KuaiRand we compared three specifications of the same logistic structure with held-out users: design slopes with re-estimated intercepts (click AUC 0.507), refitted slopes on the same proxies (0.541), and refitted slopes with viewing history and video attributes (0.688; engagement AUC 0.807).
  • Propagation: the paper uses one specification, a multivariate Hawkes process with exponential decay, a standard model for self-exciting cascades. The new snapshot data compare it with the paid schedule alone: adding propagation lowers the cumulative-curve RMSE for reads (0.100 to 0.034) and likes (0.069 to 0.031), leaves collects unchanged (0.050 to 0.048), and raises it for comments (0.047 to 0.105), whose fitted accumulation rate of 0.46 per day is the fastest of the four. The revision reports these comparisons in App. A.3–A.4.

Comments on the paper

3 comments
AuthorsOpenReview ↗

General Response to All Reviewers (1/2)

Thanks to all three reviewers. The reviews ask whether each part of OranSim is checked against observed data, and what the rollout adds to a campaign decision beyond the engagement predictor. For these questions we tested the rollout on a held-out week of a paid campaign and ran five experiments on our existing data; the revision includes all of them.

1. Held-out test on a paid campaign. We obtained the advertising reports of a paid TikTok campaign run by a consumer-electronics brand in North America: 20 accounts, 691 ad creatives, USD 769 of spend and 1.48M impressions from 2026-10-15 to 11-04, with daily spend and responses per creative and the targeting of each ad group. Our RedNote notes record no promotion spend, so the test uses this campaign. The rollout is written for a platform p (Eq. 1). The paper's code runs here with p set to TikTok and a population built from public audience statistics of the three targeted countries; targeting and spend come from the ad groups. Its response intercepts and spend-to-reach conversion are re-estimated on the first two weeks (USD 545); the slopes and propagation parameters keep the paper's values. The split, models and metrics were fixed before the held-out week (10-29 to 11-04; 313 creatives, USD 224) was read.

ModelPredicted / observed engaged viewsCreative-level absolute error / observed total
Reach × mean propensity (pooled response per dollar)2.05 [1.87, 2.28]1.05 [0.87, 1.28]
LightGBM alone (caption embeddings, product, script, account, date)1.26 [1.15, 1.39]0.34 [0.25, 0.44]
Market-response model (response per dollar by targeted market)1.09 [0.99, 1.21]0.25 [0.18, 0.33]
Rollout1.08 [0.98, 1.21]0.26 [0.19, 0.34]

Engaged views (6-second views) are the campaign's optimization objective; brackets are 95% intervals from resampling creatives. For likes and 2-second views the rollout's ratio is 0.92 and 0.94. Reach × mean propensity over-predicts by a factor of two: in the first two weeks, ad groups that also targeted Mexico returned 1,740 engaged views per dollar against 435 for US-only ad groups, and the held-out week ran US-only ad groups. The rollout draws each ad group's audience from its targeted markets, and a market-response model that makes the same distinction has the same error. The revision adds this test as §5.5, with the tables for all three outcomes in a new App. G.

2. The engagement predictor holds up over time with semantic text features. On the 39k RedNote corpus split by publication date (training 2026-03-17 to 04-13, test 04-14 to 04-20; 3,652 test notes, none of them in training), log-scale R² is 0.562 / 0.575 / 0.503 / 0.443 for reads / likes / collects / comments, and Spearman ρ is 0.671–0.714. The predictor is higher than every simple baseline on all four counts; for collects, ρ is 0.671 against 0.245 for follower count alone, 0.534 for the creator's historical median, 0.589 for metadata alone and 0.502 for text alone.

3. Table 1 with note-grouped folds. Table 1's 39,000 snapshots come from 24,542 distinct notes, and its five folds split snapshots. With folds grouped by note and PCA fitted within training folds, R² is 0.504 / 0.539 / 0.482 / 0.419 with fold standard deviations of 0.013–0.018, and the text embeddings add 0.075–0.143 over metadata alone. The note-disjoint temporal split in item 2 gives a similar range (0.443–0.575). The revision reports the grouped protocol in Table 1 and the abstract, with the row split and its fold standard deviations (0.007–0.012) beside it.

AuthorsOpenReview ↗

General Response to All Reviewers (2/2)

4. The rollout's total response has an exact closed form. M₁₄ = w·K·m̄·F, where K is reach, m̄ the mean initial response of the consumers selected by exposure ranking, w the population weight and F = 1.7575 the propagation multiplier set by the paid schedule, β, r and the time step. It reproduces all 450 rollouts (3 populations × 30 seeds × 5 options) with relative error below 6×10⁻⁸. The 1.96 ratio is therefore 2.00 (reach) × 0.981 (mean response of the enlarged audience). A model that multiplies reach by the population-average response ties every budget pair under the response-per-CNY objective; the rollout separates them through the composition of the added audience. Beyond totals, the rollout gives segment shares of response by day; the cumulative shares differ from the selected audience's composition by a total-variation distance of 0.05–0.10.

5. Three rollout components compared with observed data.

  • Population: simulated RedNote users are 81.0–81.3% female, against 80.8–83.4% in the follower profiles of 12,983 RedNote creators (total-variation distance 0.005). The simulated age distribution is older (distance 0.235); the revision reports this deviation with the profile marginals.
  • Response: on KuaiRand's randomized exposures with held-out users, re-estimating the intercepts matches the observed mean click and engagement rates (ratios 1.005 and 0.997), and the same logistic structure with viewing-history and video features reaches AUC 0.688 for clicks and 0.807 for engagement.
  • Timing: on 489 RedNote notes first observed within a day of publication, the rollout's cumulative curve is within RMSE 0.034 / 0.031 / 0.048 of the observed median for reads / likes / collects, closer than the paid schedule alone (0.100 / 0.069 / 0.050).

The revision adds a table to App. A that marks each parameter as fitted to observed notes, re-estimated on the campaign (item 1), compared with observed data, or a design value.

6. Which experiment tests which component. The revision adds this table to §4 in condensed form.

ExperimentComponentDataWhat it measures
§5.1, new temporal and grouped splitspredictor g_θ39k and 12k RedNote notesengagement prediction across notes, time, creators and niches
§5.2, new closed form and ablationsrollout R_φsynthetic populationhow each action passes through exposure, response and propagation
New paid-campaign test (§5.5)rollout R_φpaid TikTok campaign, 691 ad creativesheld-out error of predicted engaged views, likes and 2-second views
New population checkpopulation P (A.1–A.2)follower profiles of 12,983 creatorsgender and age marginals
New response fitresponse model (A.3)KuaiRand randomized exposurecalibration of the logistic structure
New timing checkpropagation (Eq. 2)RedNote note snapshotsearly accumulation of engagement
§5.3estimators for content allocation and audience selectionKuaiRand, Open Bandit, X5which estimator or learner to use when exposure or treatment logs exist
§5.4feature-substitution scoring with g_θpaired synthetic outcomeserror of predictor-based counterfactual scoring

Corrections in the revision.

  • Table 1 and the abstract use the note-grouped protocol (item 3).
  • The case's feature vector put the image count (0 for both videos) in the slot the predictor was trained on as the topic count (0.15 for both creatives). With the corrected input, predicted collects become 1,649 for s₀ and 1,781 for s_k, and the ranking of the five options is unchanged.
AuthorsOpenReview ↗

Revised Paper Uploaded

We have uploaded the revised paper. The main text is 10 pages; Section 5.5 and Appendix G are new. Section and appendix numbers of the reviewed version are unchanged, and locations refer to the revised PDF. Each change is followed by the reviewer point it answers.

Main text

  • Abstract and Section 1: the paid-campaign test in one sentence each; the abstract's R² values are the note-grouped ones (General response, items 2 and 3).
  • Section 3.2: the content channel of the rollout uses hashed vectors, and semantic vectors enter through the content-match term of Eq. 1 (RXJh Q5).
  • Section 3.5: next to Eq. 7, which decisions the predictor makes (creative, creator) and which the rollout makes (budget tier) (RXJh Q3).
  • Section 4: a table of which experiment tests which component, with the data and what is measured; Section 4.1 adds the two campaign metrics (4gDy W2 and Q3; RXJh W2).
  • Section 5.1 and Table 1: note-grouped folds as the main rows, snapshot folds alongside, fold standard deviations and interval coverage in the notes, and the note-disjoint temporal split of the 39k corpus (RXJh Q7; 2RS4 W1 and Q1).
  • Section 5.2 and Table 2: corrected predicted collects for s₀ and s_k; the closed form 1.96 = 2.00 × 0.981; the two ablations in one sentence (RXJh W1, Q2 and Q5).
  • Section 5.3: states that these estimators are for campaigns with exposure or treatment logs and do not run the rollout (4gDy W2 and Q3; RXJh W2).
  • Section 5.5 and Table 5 (new): population, response and timing compared with observed data, then the rollout on a paid TikTok campaign: calibrated on two weeks, predicting the held-out week, against reach × mean propensity, a market-response model and LightGBM only (RXJh Q1 and W2; 4gDy W1 and Q1).
  • Sections 6 and 7: Limitations now separates what has been compared with observed data from what is still a design value; the temporal-holdout count-scale error is added (RXJh Q1, Q4 and Q6).

Appendix

  • A.1 and A.2, Table A.2: sources of each population marginal and the follower-profile comparison (RXJh Q4).
  • A.3 and Table A.3: three settings of the logistic response structure on KuaiRand (RXJh Q6; 4gDy Q2).
  • A.4 and Table A.4: the closed form, its error over 450 rollouts, and the accumulation rates of 489 notes next to β and r (RXJh W1 and Q6; 4gDy Q2).
  • Table A.1: each parameter marked as fitted to observed notes, re-estimated on the campaign, compared with observed data, or a design value (RXJh Q6).
  • Table A.5: the component-by-component comparisons used in Section 5.5.
  • B.8 and Table B.9: closed form and simpler models, including reach × the whole-population mean response and the predictor with a budget term (RXJh Q2).
  • C.5, Tables C.5 and C.6: grouped folds, the temporal holdout, simple predictors, and three training variants (RXJh Q2 and Q7; 2RS4 W1 and Q1).
  • G and Table G.1 (new): campaign data, population construction, what was re-estimated, and all three outcomes at creative and ad-group-day level.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.