acceptodds
Under review as a conference paper at ICLR 2027

Curves, Not Points: Structured Curve Approximation for Engagement Dynamics in Short-Form Video and Interactive Live Streaming

Abstract

Audience engagement with video is inherently a trajectory, yet it is commonly reduced to scalar quantities such as watch time or completion rate. We instead model engagement as a structured function of time using low-dimensional parametric curve approximators. We first study live-stream engagement in KuaiLive-M3, where viewers may enter and leave throughout a room. We model occupancy as \(O=J-X\), using structured approximators for cumulative joins and exits, with a causal GRU predicting compact exit-curve parameters from evolving room representations. The resulting model achieves \(55.9%\) \(I_5\), where \(I_5\) is the fraction of rooms whose whole-trajectory occupancy IAE is at most \(5%\), compared with \(49.0%\) for the strongest independent pointwise neural baseline. We then study cross-audience retention transfer on KuaiRec-small and KuaiRec-big, which contain the same videos viewed by disjoint populations. We factor the retention representation into a video-specific parameter and behavioral watch-style parameters shared across videos. The resulting model achieves \(75.3%\) \(C_5\), where \(C_5\) is the fraction of test videos whose retention curve has sup-norm error at most \(5%\), compared with \(68.3%\) for the strongest neural baseline. Together, these results show that structured engagement curves provide compact functional representations that support accurate prediction, composition, and cross-audience transfer across distinct video-engagement regimes.

Primary area:
learning on time series and dynamical systems
Keywords:
Video engagement modeling, structured curve approximation, audience retention, live-stream occupancy prediction, sequential modeling, recurrent neural networks.

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_dRuqOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes structured curve representations for audience engagement. It combines a GRU-driven departure model with an arrival curve to predict live-room occupancy, and separates video-specific and watch-style parameters for cross-audience retention transfer. Experiments on KuaiLive-M3 and KuaiRec report gains over neural, forecasting, and survival baselines, supported by ablations on parameter sharing, residuals, audience composition, and unknown stream duration.

Critical strengths
  • The live-stream model improves on a similarly sized shared GRU. This provides meaningful evidence that the structured output helps prediction beyond reducing the size of independent pointwise models.

  • The watch-style factorization is supported by comparisons with content-based grouping and alternative parameter allocations. These experiments provide a useful insight into what enables retention transfer across audiences.

Critical weaknesses
  • The stated occupancy definition uses each viewer’s first entry and final departure. For viewers who leave and return, this counts intervening absences as occupancy. Since the appendix explicitly considers re-entry, the treatment of repeated sessions must be clarified to establish that the main experiment predicts actual occupancy.

  • The compact whole-trajectory claim needs qualification. Live-stream parameters vary with time, so three outputs at each step do not constitute one fixed three-parameter trajectory. The evaluation also appears to use progressively updated room features, leaving unclear how much the results support forecasting beyond an observed prefix.

Questions for authors
  1. Are occupancy labels constructed from every viewing session, or from each viewer’s first entry and final departure? How are intervening absences handled?

  2. For Table 1, are parameters recomputed using newly available room features at each evaluated time? Does any reported experiment freeze the observed prefix and predict the subsequent trajectory?

  3. Could the authors clarify how the 128-dimensional room features are generated, including the visual encoder and frame sampling, how visual information is combined with ASR/text and metadata, and the PCA reduction procedure?

2 comments
AuthorsOpenReview ↗

Reviewer dRuq Specific Questions (Common questions addressed on top).

We thank the reviewer for the detailed questions and constructive feedback. We address the common questions of the reviewer on instantaneous audience occupancy and fixed compact future trajectory at the top.

Construction of the 128-dimensional room features.

The 128-dimensional live-room features are not generated by us: KuaiLive-M3 ships them as timestamped segment embeddings, and the release does not include the original video, so we cannot re-encode frames or optimize the representation for occupancy. KuaiLive docs state that each feature 128 vector is a PCA reduction of an original 2,307 multimodal feature (visual, ASR/text, and related live-room signals fused before the PCA step). Public Kuaishou live-content work on the same platform describes the visual stream as a vision transformer (Swin) on sampled frames and the text stream as BERT over concatenated ASR and comments. Because we observe only the frozen 128-d PCA vectors, we cannot retune that encoder for our task; Section 4.3 (Table 5) reports the corresponding ceiling: with ideal representations the mean error would be 2.23%, versus 5.53% with the released embeddings.

We hope these clarifications and new results encourage you to raise your score.

AuthorsOpenReview ↗

Follow-up on Rebuttal and Remaining Concerns

Dear Reviewer dRuq,

We would appreciate your feedback on whether our responses and additional experiments address your concerns. If so, we hope these clarifications can be reflected in your updated assessment. If you still have reservations, it would be helpful to understand which points remain unresolved so that we can address them before discussion closes.

Thank you for your time.

Reviewer_q6YrOpenReview ↗
Rating
2
Confidence
3
Informativeness
3(+1)
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies engagement trajectories in two settings. For KuaiLive-M3, it represents audience occupancy as cumulative joins minus cumulative exits, using a shared cure-Kumaraswamy joining model and a causal GRU that predicts time-dependent parameters for the complementary exit curve. The components are jointly optimized for occupancy. For KuaiRec, it combines a video-specific Burr scale parameter with watch-style-specific shape and residual parameters, fitting the video parameter from a source-audience retention curve and transferring to a target audience using its observed group proportions. The paper reports occupancy IAE of 5.53% and I5 of 55.9% on KuaiLive-M3, and transfer IAE of 1.90% and C5 of 75.3% on KuaiRec.

Critical strengths
  1. The cross-audience factorization is a useful and testable modeling idea. Separating a video coordinate from shared behavioral coordinates gives a concrete mechanism for adapting the same video's retention curve to a different audience. The parameter-allocation and mixture-information ablations support this interpretation: Table 36 reports IAE of 6.71% when reusing the source mixture, 2.59% with a uniform mixture, and 1.90% with the full target mixture.

  2. The paper provides substantial empirical qualification of its claims. It distinguishes oracle reconstruction from held-out prediction, evaluates raw geometric violations, reports repeated training seeds, and explicitly tests clocks that do not use final room duration. The family and residual ablations also show which components contribute most.

Critical weaknesses
  1. The live-stream target is not necessarily actual audience occupancy. Section 3 defines each viewer by their FIRST entry and FINAL departure and then identifies the difference of those cumulative counts with the number of active viewers. For a viewer who leaves and later returns, this definition counts the intervening absence as continued presence. Appendix C.3 explicitly uses re-entry probability and the number of sessions as behavioral features, so the distinction needs to be resolved for this dataset. The formula is valid for a first-to-last presence envelope, or for single uninterrupted sessions, but its equivalence to actual occupancy is not established. If gaps are appreciable, the primary prediction target and all occupancy results may need to be reconstructed and reevaluated.

  2. The live-stream evaluation does not clearly establish the claimed compact future-trajectory prediction. The displayed model evaluates the curve at time u using parameters produced from representations available through that same u. Since those parameters vary with u, the resulting trajectory is not necessarily described by one fixed three-coordinate vector; it may instead be a causal, pointwise prediction with a structured output transformation. No explicit observation cutoff and future-only evaluation horizon are specified. The primary experiment also uses final duration, and the paper appropriately acknowledges this: Table 6 shows IAE worsening from about 5.5% to 7.5% without that information. However, the corresponding online baseline comparison is not reported. These issues limit both the forecasting interpretation and the evidence that the proposed structure improves genuinely prospective prediction.

Questions for authors
  1. Are repeated viewing sessions collapsed into first-entry/final-exit intervals when constructing the labels? What fraction of viewers re-enter, and how much do the resulting curves differ from occupancy reconstructed from the union of actual session intervals? If the difference is material, do the main rankings persist on session-correct occupancy?

  2. At a fixed observation cutoff, are the predicted exit parameters frozen to generate the future curve, or recomputed using embeddings from each evaluated time? Please specify the information available at prediction time, including embedding timestamps and duration information, and report a future-only, unknown-final-duration comparison against the strongest shared-GRU baseline. A contemporaneous trajectory-estimation task would also be legitimate, but would require narrower claims.

  3. For KuaiRec, do all competing predictors receive the same full source-audience curve and the same source and target watch-style proportions, with the same restriction on held-out target outcomes? Please specify these inputs and the adaptation of DeepTime, DSM, and Dynamic-DeepHit, including any range or monotonicity postprocessing. This is important because the observed audience composition accounts for a large part of the transfer gain, and the event-time baselines perform much worse than even simple direct predictors.

2 comments
AuthorsOpenReview ↗

Reviewer q6Yr Specific Questions (common questions addressed on top).

We thank the reviewer for the detailed questions and constructive feedback. We answer the reviewers questions on instantaneous audience occupancy, fixed compact future trajectory and online baseline comparisons in the common questions section at the top. Here we answer the reviewers question on the information availabe to the KuaiRec predictors and the fairness and adaptation of the time-series methods.

Information available to KuaiRec predictors and fairness of the baseline comparison.

There was an minor error in how watch-style proportions were passed to the time-series methods which we fixed. A bigger issue was on the pointwise and along-u nets. Those architectures were sized for KuaiLive, which has enough samples to support them. KuaiRec is much smaller (~2300 training videos). With the original weight counts — for example 275k for pointwise MLP and 416k for pointwise GRU, versus k=1 for Burr — those nets were memorizing the dataset rather than learning, so the scores they posted were higher than a fair comparison should allow. We dropped the neural weight budgets to about the size of the KuaiRec training set (k ≈ 2000) and and reran all experiments (including time-series methods). Yes, all competing predictors receive the same full source-audience curve and the same source and target watch-style proportions, with the same restriction on held-out target outcomes.

FamilyModelIAEC5C10kk
pointwiseridge_per_bin2.16%63.4%95.1%500
pointwisemlp_per_bin24.64%0.0%0.0%1936
pointwisegru_per_bin28.16%0.0%0.0%1850
pointwisernn_per_bin36.41%0.0%0.0%1950
along-uumlp_along_u4.12%2.6%71.7%1961
along-uugru_along_u6.56%0.0%0.0%1881
along-uurnn_along_u3.95%0.0%71.7%1990
mix-head pointwisemlp_per_bin24.37%0.0%0.0%1989
mix-head pointwisegru_per_bin26.33%0.0%0.0%1859
mix-head along-uumlp_along_u5.96%0.9%48.2%1963
mix-head along-uugru_along_u20.94%0.0%0.0%2000
time seriesdeeptime13.25%7.1%22.6%6922
time seriesdsm14.46%0.0%0.0%4936
time seriesddh27.29%0.0%0.0%42371
structured mixBurr+G20+sine31.90%75.3%98.7%1

The original pointwise / along-u numbers were the ones that were too optimistic: those nets were running under favorable conditions (far more weights than KuaiRec can support), so they looked closer to Burr than they should have. Once capacity is matched to the dataset, the structured curve remains the clear favorite (IAE 0.0190, C5 75.3%). We next describe the adaptiations of the time series algorithms to our domain.

DeepTime. A d=10 dictionary g_φ(τ) is fit so that a linear readout of g_φ matches Ŝ_big on fit videos. The readout W ∈ R^{5×10} is a soft-grouped ridge on fit-video Ŝ_small only: video v contributes to class c with weight π_small,v,c. At test we mix with the target composition, ŷ(τ) = g_φ(τ)^T (π_big W). Test Ŝ_small and Ŝ_big never enter W. Predictions are clipped to [0,1]; there is no extra monotonicity projection.

DSM. Covariate is x = [π_small, π_big]. For each fit video we form quantile hitting times from Ŝ_small only: T_q = inf{u : Ŝ_small(u) ≤ q} at q = 0.9, …, 0.1, right-censored at u = 1. DSM is a 4-Weibull mixture for P(T_q > t | x). The knot is E[T_q | x], clipped to [0,1], then made isotonic (T_0.9 ≤ ⋯ ≤ T_0.1). The predicted curve is the linear interpolant through (0,1), (E[T_0.9], 0.9), …, (E[T_0.1], 0.1), clipped to [0,1]. Test Ŝ never enters the likelihood.

Dynamic-DeepHit. Same T_q construction and the same x = [π_small, π_big]. One visit at u = 0 (the remaining-audience curve is the target, so it is not a longitudinal input). Shared GRU+attention encoder, nine softmax heads over M=50 time bins, knot E[T_q | x] = ∑m τ_m o{q,m}. Same isotonic hitting times, same interpolant, same clip to [0,1]. Test series never enter the DDH losses.

We hope these clarifications and new results encourage you to raise your score.

AuthorsOpenReview ↗

Follow-up on Rebuttal and Remaining Concerns

Dear Reviewer q6Yr,

We would appreciate your feedback on whether our responses and additional experiments address your concerns. If so, we hope these clarifications can be reflected in your updated assessment. If you still have reservations, it would be helpful to understand which points remain unresolved so that we can address them before discussion closes.

Thank you for your time.

Reviewer_wPNEOpenReview ↗
Rating
2
Confidence
3
Informativeness
3(+1)
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a structured curve approximation approach for modeling video engagement dynamics. It aims to address the limitations of existing studies that typically reduce user viewing behavior to scalar metrics, such as watch time or completion rate, thereby failing to capture the complete temporal dynamics of user engagement.

Critical strengths
  1. Motivation is reasonable and practically meaningful. Modeling user engagement dynamics through complete temporal curves, rather than relying solely on scalar metrics, is a meaningful and promising research direction.

  2. The structured curve modeling approach offers a certain degree of interpretability. Compared with independently predicting each time point, the proposed method represents the entire engagement trajectory using a small number of parameters, which helps reduce the dimensionality of the output space and introduces structural constraints into the predictions.

  3. Experimental evaluation is relatively comprehensive, providing evidence to support the effectiveness of the proposed approach.

Critical weaknesses
  1. The dependence on the final live-stream duration limits the practical applicability of the method in online settings. In the KuaiLive-M3 setting, the authors use normalized time (u = t/L) to model audience occupancy dynamics, where (L) denotes the final duration of the live stream. However, in real-world online live-streaming scenarios, the final duration is typically unknown in advance. Although the authors discuss this limitation in Section 4.4, I suggest that they further elaborate on how the proposed method could be deployed in practical online settings, for example, by jointly predicting the live-stream duration and audience engagement trajectory, or by developing an online curve prediction mechanism that does not require knowledge of the final duration. Furthermore, given the practical and industry-oriented nature of this work, including online A/B testing results would make the empirical evaluation more convincing.

  2. The methodological connection between the two tasks is relatively weak, and the overall methodological novelty could be further strengthened. Although the paper presents Structured Curve Approximation as a unified research direction, the KuaiLive-M3 and KuaiRec tasks employ substantially different modeling approaches. Moreover, considering that parametric curve fitting and related modeling techniques have already been extensively studied, I suggest that the authors more clearly articulate the key technical novelty of the proposed approach compared with existing methods and provide a more thorough explanation of the unified methodological contributions across the two tasks.

  3. The dependence on specific curve families and parameterization choices deserves further discussion. The paper selects different parametric models for the two datasets by exploring various curve families, parameter-sharing strategies, and residual structures. Although the authors provide relatively comprehensive model selection experiments and ablation studies, the final approaches still involve several design choices tailored to specific data distributions. It would be even better if the author could further explore how this structured curve modeling framework can be extended to other video platforms or datasets with different viewing behavior distributions, in order to provide stronger evidence of generalization ability.

If my concerns can be properly addressed, I will improve my scores.

Questions for authors

Please refer to Weaknesses.

4 comments
AuthorsOpenReview ↗

Reviewer wPNE Specific Questions (common questions addressed on top).

We thank the reviewer for the detailed questions and constructive feedback. (We answer the common question on online algorithm at the top)

1. Methodological connection between KuaiLive-M3 and KuaiRec, and overall novelty.

The common contribution is to exploit engagement-curve geometry for prediction, even when the desired output is retention at a particular time(s). The key novelty is to use curve templates whose coordinates are stored by neural networks learnt from available observations to predict retention. Additionally, we also show that learning one curve per watch-style-specific class enables robustness in audience transfer settings.

KuaiLive-M3. Our major evaluation dataset is KuaiLive-M3 which has timestamped room representations along with retention data to experimentally validate our key novelty. We show that using curve templates whose coordinates are stored in a GRU, outperforms all other baselines in predicting retention. We further do investigate the effect of learning one-curve per watch-style specific class for this datastet - however in the absence of audience-transfer data for this dataset, we must contend with evaluating this approach on the same audience data (see Appendix C.3 , Tables 13-14) which shows no improvmeent over the watch-style agnostic algorithm on the same audience.

KuaiRec. We wanted to repeat the same experiment on another dataset to show robustness - however this requires the dataset to contain both timestamped content embeddings AND proprietary retention labels which most major datasets like KuaiRec, KuaiRand, Tenrec, YouTube-8M etc lack. KuaiRec nevertheless provides retention data on the same video for different viewing audiences (without giving timestamped content embeddings) which allows us to validate robustness of using structured curves to predict retention in audience transfer settings. In the absence of timestamped content embeddings, we optimize the video-specific coordinates directly from the source-audience retention curve, approximating an ideal GRU predictor whose outputs are unrestricted by a learned mapping from content to coordinates.

Thus, the two tasks share a structured curve representation but differ based on the information given by the dataset. We use KuaiLive to validate the key novelty -- that structured curves with GRU parameters predict retention best wheras KuaiRec allows validates robustness of structurved curves to predict retention on audience transfer settings. Finally, the selected curve templates differ because the datasets capture distinct video distributions: KuaiLive-M3 contains long, interactive live streams of 3.42 hours on average, whereas KuaiRec contains short-form videos of 10 seconds on average.

2. Dependence on selected curve families and generalization to different data distributions.

The selected curve families and parameterizations are dataset-specific instances of a general modeling procedure. Applying the framework to a new dataset does not require assuming that the same curve family will remain optimal. Given training, validation, and held-out test sets, the procedure is:

  1. Select a base template. Train GRU coordinate predictors for candidate curves from our approximator library using training data, and compare their prediction errors on validation data.

  2. Evaluate residual extensions. For the selected base predictor J^\hat J, model the training residual J−J^J-\hat J using candidate damped sine/cosine terms and Gaussian bumps. Then jointly train each base-plus-residual model and compare validation errors. Retain the strongest residual extensions and the base-only model for next steps - often these residual extensions overfit rather than learn.

  3. Select parameter sharing. Compare allocations in which each coordinate is either shared across videos or predicted separately for each video by the GRU. Select the allocation on validation data. Appendix C.7, Tables 30–32 illustrate this allocation analysis for KuaiRec, where video-specific coordinates are fitted from source curves rather than predicted from content embeddings.

  4. Introduce audience structure where useful. Construct watch-style clusters from training viewing behavior, such as early skipping, completion, normalized watch duration, and rewatching. Demographic attributes can provide additional covariates, but do not themselves define behavioral watch style. Then train a separate GRU coordinate predictor for each cluster, using the selected template and parameter allocation, and combine the resulting curves using the target audience's cluster proportions. The clustering rule and predictors must be fixed before test evaluation.

This procedure provides a general approach for adapting structured curve prediction to different data distributions. KuaiLive-M3 evaluates GRU-based coordinate prediction, including watch-style-specific models, while KuaiRec evaluates the transfer of structured curves across audiences.

AuthorsOpenReview ↗

Comment part 2.

We hope these clarifications and new results encourage you to raise your score.

AuthorsOpenReview ↗

Follow-up on Rebuttal and Remaining Concerns

Dear Reviewer wPNE,

We would appreciate your feedback on whether our responses and additional experiments address your concerns. If so, we hope these clarifications can be reflected in your updated assessment. If you still have reservations, it would be helpful to understand which points remain unresolved so that we can address them before discussion closes.

Thank you for your time.

Thank you for your detailed response. After referring to the comments from other reviewers, I believe the quality of this paper does not meet the "weak accept" standard. Additionally, this paper has conducted many experiments, but I think it would be more convincing if it were deployed in a large-scale scenario. I prefer to consider this as a "boardline" paper. I have raised the score for the "Informativeness" sub-socre. If other reviewers are willing to support the acceptance of this paper, I will not object.

Comments on the paper

3 comments
AuthorsOpenReview ↗

Common Questions.

We thank all reviewers for their careful reading and constructive feedback. Here we address the common questions raised across reviews. We address each reviewer's remaining specific questions in separate replies to their reviews.

Common Question 1: Dependence on live-stream duration and online baseline comparisons. [Reviewers wPNE, q6Yr]

Section 4.4 in the paper shows that replacing normalized progress with absolute-time or logarithmic clocks reduces the accuracy of our structured model. We now extend this evaluation with duration-unknown baselines and experiments using an imperfect planned duration.

For online deployment, the creator must pre-determine an intended duration or expected end time before streaming. This provides a planned duration L′L' for normalization, while allowing the actual duration LL to differ upto a wide margin for error (specifically (−15(-15 \\% , + 23\\% ] error is allowed). We show by experiments that this approach is superior to all other baselines when adapted to the online setting, and is only inferior to our offline approach. We also give results showing the performance drop vs error growth. Creator intentions and external scheduling constraints need not be encoded in the observed video, so predicting duration from video features is not a prerequisite for this deployment approach. The KuaiLive dataset contain completed recordings whose duration lengths are known; and hence does not represent a truly online setup. Our experiments therefore evaluate online information constraints retrospectively, wheras as mentioned earlier predicting duration is not a pre-requisite for this approach.

Sensitivity to planned duration. We first show how the O IAE, O sup and I5 errors vary with k=(L′−L)/L∗100k = (L'-L)/L * 100.

kkO IAEO supI5
−30-30\\%8.048.04\\%19.1219.12\\%23.823.8\\%
−25-25\\%7.457.45\\%18.0118.01\\%29.729.7\\%
−20-20\\%6.906.90\\%16.8516.85\\%35.735.7\\%
−15-15\\%6.406.40\\%15.7215.72\\%42.142.1\\%
−10-10\\%5.985.98\\%14.7214.72\\%49.049.0\\%
−5-5\\%5.665.66\\%14.0014.00\\%53.953.9\\%
00\\%5.535.53\\%13.6813.68\\%55.955.9\\%
+5+5\\%5.665.66\\%14.0914.09\\%53.553.5\\%
+10+10\\%5.875.87\\%14.8114.81\\%50.050.0\\%
+15+15\\%6.096.09\\%15.6615.66\\%46.546.5\\%
+20+20\\%6.296.29\\%16.4816.48\\%43.343.3\\%
+25+25\\%6.466.46\\%17.2317.23\\%40.240.2\\%
+30+30\\%6.626.62\\%17.9017.90\\%37.837.8\\%

Comparison without duration input. We next evaluate our above-mentioned online structured curve approach and compare it with other baselines when adapted to an online setting: both for an absolute clock, or for a lograrithmic clock (along-u neural architecture approaches do not use a clock).

FamilyMethodClockOnline?O IAEO supI5
proposedjoint Ot/Lt/Lno5.535.53\\%13.6813.68\\%55.955.9\\%
proposedjoint O (online)t/L′t/L', k=5k=5\\%yes5.665.66\\%14.0014.00\\%53.953.9\\%
proposedjoint O (online)t/L′t/L', k=10k=10\\%yes5.875.87\\%14.8114.81\\%50.050.0\\%
proposedjoint O (online)t/L′t/L', k=15k=15\\%yes6.096.09\\%15.6615.66\\%46.546.5\\%
proposedjoint O (online)t/L′t/L', k=30k=30 \\%yes6.626.62\\%17.9017.90\\%37.837.8\\%
pointwiseMLP N=10N=10absyes15.2115.21\\%25.6825.68\\%12.512.5\\%
pointwiseRNN N=10N=10absyes15.1815.18\\%25.4125.41\\%13.013.0\\%
pointwiseGRU N=10N=10absyes15.4415.44\\%25.5625.56\\%13.413.4\\%
pointwiseGRU N=100N=100absyes9.529.52\\%20.5820.58\\%24.224.2\\%
pointwiseMLP N=10N=10logyes16.2916.29\\%26.3026.30\\%12.212.2\\%
pointwiseRNN N=10N=10logyes9.689.68\\%20.2120.21\\%23.223.2\\%
pointwiseGRU N=10N=10logyes9.289.28\\%19.9419.94\\%26.226.2\\%
pointwiseGRU N=100N=100logyes7.237.23\\%20.3820.38\\%32.432.4\\%
along-uMLPNAyes16.2616.26\\%26.0826.08\\%12.312.3\\%
along-uRNNNAyes6.806.80\\%19.2819.28\\%33.533.5\\%
along-uGRUNAyes6.396.39\\%18.1718.17\\%41.441.4\\%
time-seriesDeepTimeabsyes8.658.65\\%18.5018.50\\%24.524.5\\%
time-seriesDeepTimelogyes8.648.64\\%18.5218.52\\%24.924.9\\%
time-seriesDSMabsyes9.579.57\\%19.0019.00\\%11.811.8\\%
time-seriesDSMlogyes10.5210.52\\%25.5925.59\\%7.47.4\\%
time-seriesDynamic-DeepHitabsyes9.569.56\\%18.9618.96\\%11.711.7\\%
time-seriesDynamic-DeepHitlogyes12.1612.16\\%25.3725.37\\%2.22.2\\%

These results support deployment with an approximate planned duration as the best approach for the online setting.

AuthorsOpenReview ↗

Additional Common Questions.

Common Question 2. Does the target represent actual instantaneous occupancy when viewers re-enter? [Reviewers q6Yr, dRuq]

Our target intentionally uses each viewer's first entry and final departure and does not model temporary departures followed by re-entry. This choice is intentional. A viewer's final departure can depend on the room and on the content/story experienced up to that point: if the content is not sufficiently interesting, the viewer can leave and not return. In contrast, a temporary departure followed by later re-entry can depend on exogenous factors not represented by the video-content history, such as a phone call, connectivity interruption, power loss, or another offline task. Modeling every intermediate departure as a content-driven departure would therefore conflate durable disengagement with temporary external interruptions.

Accordingly, our intended target is first-entry-to-final-departure engagement, rather than session-union occupancy that treats every intermediate absence as a departure. Furthermore, re-entry probability and number of sessions are not inputs to our main KuaiLive-M3 method. They appear only in Appendix C.3, where we construct optional watch-style clusters for an ablation of class structure to show accounting for such class structures yields no improvement to the main algorithm.

For completeness, we retrained every Table 6 baseline on session-union occupancy (differs from envelope occupancy by 12.2% as measured by L1 norm) and the coarse ranking among baselines is unchanged - along-u GRU obtains slightly better average error than structured curves, but is worse in max error and I5. Surprisingly, train-and-test IAE on session-union occupancy is lower than on the first-to-last envelope: A possible explanation is that accounting for intermediate absences lowers occupancy levels and the absolute magnitude of their fluctuations, thereby reducing absolute prediction error.

FamilyMethodO IAEO supI5
proposedjoint O2.512.51\\%7.657.65\\%90.790.7\\%
pointwiseMLP N=10N=104.084.08\\%10.1810.18\\%73.473.4\\%
pointwiseRNN N=10N=103.013.01\\%9.359.35\\%84.484.4\\%
pointwiseGRU N=10N=103.233.23\\%9.749.74\\%80.780.7\\%
pointwiseGRU N=100N=1003.843.84\\%10.9510.95\\%75.075.0\\%
pointwiseGRU N=500N=5003.603.60\\%11.7511.75\\%77.177.1\\%
along-uMLP4.074.07\\%10.0610.06\\%73.473.4\\%
along-uRNN2.572.57\\%8.788.78\\%90.690.6\\%
along-uGRU2.452.45\\%8.158.15\\%90.190.1\\%
time-seriesDeepTime3.503.50\\%8.538.53\\%82.182.1\\%
time-seriesDSM5.095.09\\%11.3511.35\\%62.262.2\\%
time-seriesDynamic-DeepHit5.125.12\\%11.5911.59\\%61.861.8\\%

Common Question 3. Does the model establish a fixed compact future trajectory? [Reviewers q6Yr, dRuq]

We agree that our usages of the phrases “future trajectory prediction” and “one fixed three-coordinate trajectory” can suggest a different frozen-prefix task; we will revise these terms. Our model instead uses a dynamically updated three-coordinate structured output at each time, allowing the estimate to incorporate newly observed content while retaining a constrained low-dimensional output interface. Specifically, the template coordinates may vary with time: ψr(u)=gϕ(er,≤u)+b.\psi_r(u)=g_{\phi}(e_{r,\leq u})+b. However, the trained GRU weights ϕ\phi and bias bb remain fixed during evaluation. The changing coordinates are outputs of the same trained model where the input at time uu is the time-stamped R128R^{128}-dimension video representations for time ≤u\le u.

This distinction is central to the task. Whether a viewer leaves at say hour 10 cannot be predicted based on video representations upto a fixed cutoff for eg upto hour 3 as the viewer may choose to leave based on developments in the video after hour 3. To estimate engagement in response to the content actually experienced upto hour 10, the model must be allowed to incorporate that content. Our task instead estimates engagement based on the video content consumed so far. A frozen-prefix forecast must effectively anticipate how viewers will judge the quality of unknown future content, which is beyond the scope of our task.

Dear Reviewers,

As the discussion phase is coming to an end, could you please take a moment to review the authors' responses and finalize your ratings?

If you have any remaining concerns or questions, please feel free to raise them with the authors before the discussion closes.

Thank you very much for your time and support!

Best regards,

Xiaojuan

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.