acceptodds
Under review as a conference paper at ICLR 2027

Popularity Predictability Depends on Information Scale

Abstract

Micro-video popularity prediction draws on information ranging from broad context to item content and historical references. However, existing methods largely ignore that information at different levels contributes differently to predictability, and instead collapse them into a single prediction. We propose Scale-Structured Predictability Modeling (SSPM), which treats popularity predictability as a function of information scale. SSPM organizes increasingly informative evidence into a nested prediction process, where each level explains only the variation that becomes predictable beyond the preceding one. (1) Context Prior Modeling establishes the predictable component supported by population or temporal context. (2) Multimodal Evidence Refinement identifies the additional variation explained by item-level multimodal evidence. (3) Historical Analogy Refinement further captures recurrent local variation shared with similar historical instances, using out-of-fold residuals to separate genuinely new information from what earlier levels have already explained. Together, these components form a progressive decomposition of predictability rather than three independent predictors. Experiments on three micro-video popularity benchmarks show that SSPM improves prediction performance by up to 19.77% in nMSE while revealing how predictive gains emerge across information scales. These results support our main finding: popularity predictability depends on information scale, and explicitly modeling this structure leads to both stronger prediction and better understanding of where predictive gains come from.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
micro-video popularity prediction, information scale, multimodal learning, residual learning, historical analogy, cross-fitting, predictability attribution, retrieval-augmented prediction

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 2 reviews, above 27% of the others.
Reviewer_st2JOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies micro-video popularity prediction from the perspective of information scale. Instead of directly fusing heterogeneous signals into a single predictor, the authors argue that different information sources contribute to predictability at different statistical scopes. Based on this view, they propose Scale-Structured Predictability Modeling (SSPM), which organizes prediction into three progressively refined stages: Context Prior Modeling, Multimodal Evidence Refinement, and Historical Analogy Refinement. Each stage is trained to explain the residual variation left unresolved by the previous stage, rather than independently fitting the full popularity target.

The framework is further formalized using a nested filtration over prediction-time information. Under squared loss, the corresponding scale-wise prediction increments are interpreted as orthogonal components of predictability, which provides the statistical motivation for the residual-learning formulation.

Empirically, SSPM is evaluated on SMPD, MicroLens, and Informs, where it consistently outperforms the strongest reported baseline across nMSE, MAE, and Spearman correlation. The paper also includes matched ablations and scale-wise attribution analyses to show that the gains arise from different information sources across datasets, with multimodal evidence dominating on SMPD and MicroLens, while contextual information contributes most strongly on Informs. Overall, the paper proposes a conceptually distinct way to formulate popularity prediction: rather than treating heterogeneous features as a single fused representation, it decomposes predictability according to the scale at which additional information becomes useful.

Critical strengths
  1. The core idea is intuitive and conceptually well motivated.
    I find the proposed information-scale perspective very natural and reasonable, and I do not have major conceptual concerns about the formulation. Modeling popularity prediction as a progressive refinement process—from contextual priors, to multimodal evidence, and finally to historical analogies—provides a clear interpretation of how heterogeneous information contributes to predictability.

  2. The overall design is elegant and internally coherent.
    Rather than treating the three stages as independent predictors, the authors explicitly train each stage to model the unresolved residual left by the previous one. The use of cross-fitted residuals and residual-based historical retrieval is particularly thoughtful, as it helps prevent later stages from simply relearning information already captured at coarser scales.

  3. The experimental design provides meaningful support for the proposed principles.
    The matched ablations compare one-shot prediction, full-target learning, removal of OOF construction, direct label retrieval, and alternative neighbor-selection strategies under controlled settings. These experiments make the empirical evidence substantially more convincing than a conventional component-removal ablation.

  4. The paper provides useful insight beyond simply improving SOTA performance.
    The scale-wise attribution analysis reveals that different datasets rely on substantially different predictive sources—for example, multimodal evidence dominates on SMPD and MicroLens, whereas contextual information dominates on Informs. I find this analysis useful because it gives practitioners a clearer understanding of where predictive gains actually originate.

Overall, I find the proposed design thoughtful, coherent, and potentially useful beyond the specific popularity-prediction setting, particularly as a general way to organize heterogeneous information sources according to the additional variation they can explain.

I suggest that the authors consider citing and discussing the following two related works:

  • Lee et al., “Anchoring Trends: Mitigating Social Media Popularity Prediction Drift via Feature Clustering and Expansion,” ACM Multimedia 2025. This work studies feature clustering on SMPD to mitigate popularity prediction drift, which is relevant to the present paper’s discussion of structured information and different sources of predictability.

  • Hsu et al., “Revisiting Vision-Language Features Adaptation and Inconsistency for Social Media Popularity Prediction,” ACM Multimedia 2024. This work systematically analyzes how different vision-language features affect popularity prediction on SMPD, which is also closely related to the paper’s motivation that different information sources contribute differently to predictability.

Both papers provide relevant perspectives on how heterogeneous features and their structure influence popularity prediction, particularly on SMPD. I therefore believe that discussing them would help better position the proposed information-scale perspective with respect to closely related prior work.

Critical weaknesses
  1. The main concern is the lack of sufficient clarity regarding data preprocessing and feature construction.
    Social-media multimodal datasets are substantially more complicated than standard image or video benchmarks. Taking SMPD as an example, the raw information may involve user identity, post identity, geographic information, posting time, user statistics, images/videos, textual descriptions, post categories, interactions from other users, view statistics, follower/following counts, hashtags, and various categorical variables. Historical posts from the same user may also influence the popularity of a newly published post. A substantial body of prior work on SMPD has specifically investigated how these heterogeneous signals should be processed and integrated (e.g. HyFea). Although Appendix A.2 lists several major feature groups used by SSPM, such as the 175-dimensional base representation, text embeddings, metadata, and prior-history summaries, the paper does not provide enough detail about how these features are individually cleaned, encoded, normalized, and combined to make the experimental pipeline sufficiently transparent. This is my primary concern with the current submission.

  2. Reproducibility is particularly important for this task.
    In the popularity-prediction literature, there have been cases where strong reported results were difficult to reproduce because code was unavailable, incomplete, or did not fully correspond to the reported experimental pipeline. This makes implementation transparency particularly important for the present work. The manuscript states that an anonymized code package will be released for review, but I would like the authors to clarify whether the complete preprocessing, feature construction, training, retrieval, and evaluation pipeline will be included, and whether they intend to publicly release the full code after publication. Given the complexity of the datasets, releasing only the core model implementation would not be sufficient for reliable reproduction.

  3. The construction of the multi-scale relationship is conceptually clear but operationally under-specified.
    The paper clearly defines the hierarchy of Context Prior, Multimodal Evidence, and Historical Analogy at the methodological level, and explains that later stages learn cross-fitted unresolved targets. However, from an implementation perspective, it is still difficult to understand exactly how the heterogeneous features interact at each stage and how the cross-fitting pipeline is instantiated for each dataset. Figure 2 provides a useful conceptual overview, but it is not sufficient for reproduction by itself. 39798_Popularity_Predictabilit I strongly recommend providing a dataset-by-dataset description of the complete feature-processing pipeline, including the selected visual/text encoders, preprocessing and normalization procedures, categorical-feature handling, text cleaning, and how each feature is assigned to a particular information scale. This is especially important for SMPD, where multilingual text, emojis, hashtags, and noisy metadata can materially affect performance.

  4. The clustering component shown in the framework requires substantially more explanation.
    The clustering or topic-grouping operation illustrated in Figure 2 is not described with enough detail for the reader to understand its exact construction, design rationale, or contribution. The authors should specify the clustering features, algorithm, number or selection of clusters, fitting protocol, whether clustering is constructed using training data only, and how the resulting groups enter the Context or subsequent prediction stages. I would also like to see an ablation or sensitivity analysis demonstrating how this clustering design affects the final prediction performance and scale-wise attribution. Without this information, it is difficult to determine whether the reported gains arise from the proposed scale-structured principle itself or from particular feature-engineering and grouping choices.

Questions for authors

Please provide a comprehensive response addressing all of the weaknesses and reproducibility concerns raised above.

8 comments
AuthorsOpenReview ↗

Author Response to st2J

Thank you for the constructive feedback and references. Both recommended works are excellent and highly relevant to our study. We will discuss Lee et al. (2025) and Hsu et al. (2024) and clarify how SSPM differs from feature clustering and vision-language feature adaptation.

[W1] Data Preprocessing and Feature Construction

Re: SSPM uses fixed feature representations across stages and matched controls; differences come from residual targets and information access, not stage-specific preprocessing.

  • SMPD: Context uses historical popularity, calendar, and topic statistics; Multimodal uses visual, text, and metadata representations, including 175-dimensional benchmark features.
  • MicroLens: CLIP frame features are followed by train-fitted PCA128, with seven metadata features.
  • Informs-Strict: Visual features use train-fitted PCA64 and training-only target normalization.

Train-fitted transformations are applied unchanged to validation and test data. We will provide dataset-specific feature definitions and processing details in the appendix.

[W2] Reproducibility

Re: We will release the complete cross-fitted pipeline:

  • Preprocessing and feature construction.
  • Fold generation and model training.
  • Residual-bank construction and retrieval.
  • Evaluation scripts.
  • Model configurations, feature assignments, data splits, and historical-reference eligibility rules.

This supports reproduction beyond the core model implementation.

[W3] Operational Multi-scale Construction

Re: The stages are connected by OOF residual targets, not unrestricted feature fusion:

  1. Context: learns from population and temporal information.
  2. Multimodal: learns the Context OOF residual using current-item evidence.
  3. Historical: retrieves eligible OOF residuals from the complete preceding prefix.

Earlier predictors remain frozen during later refinement. SMPD and Informs use chronological OOF construction; MicroLens uses its declared fold protocol. This separates stage-specific inputs and measures each refinement through held-out prefix-risk reduction.

[W4] Clustering and Topic Grouping

Re: Clustering builds dynamic topic-level priors for Context; it is not an independent predictor.

  1. Construction: Using publication-time text and metadata, we evaluate MiniBatchKMeans, SVD-KMeans, tag-based MiniBatchKMeans, and NMF with K ∈ {8,10,12,16,20,30}. The selected model is NMF with 20 topics and Top-2 soft assignments, based on validation performance, coherence, silhouette score, and cluster balance.
  2. Integration: Topic weights construct historical popularity, temporal trend, and content-supply features. Topic models are fitted on training data only; historical statistics use only samples preceding each query.
  3. Ablation: The recovered SMPD A3/A6 control uses the same Ridge estimator and splits; A6 only adds topic-history features. These features reduce validation MAE by 3.66%, validation MSE by 4.38%, and test MAE by 1.91%.

This shows that topic-derived features improve the Context prefix without increasing model capacity. We will clarify the grouping mechanism in Figure 2 and the appendix.

SMPDA3: Without topicA6: With topic
Validation MAE ↓2.41232.3240
Validation nMSE ↓1.03270.9875
Test MAE ↓2.20042.1585

Thank you for the response, and sorry for the late reply. Overall, I am not yet satisfied with it. I would like the authors to clarify two points:

  1. Encoder choice. Which visual/video encoder is used, e.g., CLIP (which variant) or a video encoder such as X-CLIP? Which text encoder is used? Are the same encoders used across all three datasets? If not, what is the rationale? The strength of the representation directly affects the Multimodal stage’s share, so this bears on the attribution claims, not only on reproducibility.

  2. Feature scope. Why were broader features not considered, such as richer user statistics, audio, or VLM-generated descriptions? Do SSPM’s gains and attribution pattern still hold with stronger features?

AuthorsOpenReview ↗

Author Response to Reviewer st2j's Follow-up Questions

Thank you for your thoughtful and detailed follow-up. We sincerely appreciate your careful examination of our methodology and experimental design.

Q1. Encoder Choice and Fairness

We clarify that SSPM uses the same pretrained visual encoder, OpenAI CLIP ViT-L/14, across all three datasets. We do not use X-CLIP or different primary visual backbones across datasets.

(1) Consistent encoders. Our visual encoder follows the CLIP ViT-L/14 setting used in STAP (C4). For text, SMPD uses UAE-Large-V1, consistent with the UAE-Large encoder family reported by STAP. The frozen MicroLens and Informs pipelines do not introduce a separate pretrained text encoder.

(2) Fair comparisons. All matched SSPM controls within each dataset use the same encoder representations. The visual backbone is fixed. Only dataset-specific processing differs:

  • SMPD: 100 frames, mean pooling, PCA32.
  • MicroLens: 32 frames, mean pooling, PCA128.
  • Informs: 100 frames, mean pooling, PCA64.

All PCA transformations are fitted on training data only. These differences reflect dataset-specific feature processing, not different pretrained visual encoders.

(3) Attribution validity. Since the encoders are fixed within each dataset, the improvements in our matched controls cannot be attributed to switching visual backbones.

SSPM measures held-out risk reductions under fixed representations. The attribution shares are configuration-dependent, rather than assumed to be invariant to representation strength.

We will explicitly document these encoder configurations and feature-processing details in the revised manuscript.


Q2. Feature Scope and Robustness

SSPM already includes pre-publication user-history statistics, temporal features, and topic-level priors on SMPD.

We further evaluated stronger video representations on MicroLens. We extended Historical retrieval from CLIP to CLIP + VideoMAE. Both settings use the same 1,974 validation samples, frozen Global predictions, OOF residual targets, and retrieval budget.

Historical FeaturesMAE ↓nMSE ↓SRC ↑
CLIP1.01440.64650.5870
CLIP + VideoMAE1.01280.64350.5897

The richer representation improves all three metrics. With residual clipping, our final configuration reaches an nMSE of 0.6431. Historical refinement reduces MSE by 0.0400 over the frozen Global prefix (95% bootstrap CI [0.0213, 0.0592]).

Richer features may further improve prediction, but SSPM's distinct contribution lies in scale-wise residual learning and held-out risk attribution, rather than any specific feature combination.

Our attribution measures held-out risk reductions under each feature configuration. It therefore captures changes in predictive contributions rather than assuming fixed attribution percentages.

We will clarify the feature scope and add this controlled comparison to the appendix.

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

This paper proposes SSPM, which organizes popularity prediction into context prior modeling, multimodal residual refinement, and historical residual retrieval. Cross-fitting constructs out-of-fold training targets, while held-out risk reductions quantify improvements across successive prediction stages. Experiments cover three popularity prediction benchmarks.

Critical strengths
  1. Retaining intermediate predictions makes it possible to examine where predictive improvements occur, providing more informative diagnostics than endpoint accuracy alone.

  2. The fold-local cross-fitting procedure avoids using held-out labels to construct earlier-stage training targets. Retrieving out-of-fold residuals is a sensible approach to correcting errors on unseen instances.

  3. The paper includes matched design controls, retrieval ablations, and synthetic verification. The paired-bootstrap interval in Appendix C.1 also provides concrete evidence for the historical refinement gain on MicroLens.

Critical weaknesses
  1. The methodological contribution needs clearer positioning against closely related residual-retrieval work. ResMem already combines a base predictor with nearest-neighbor regression on its residuals. SSPM does not discuss this directly relevant approach. Its cross-fitting and explicit scale decomposition offer potential distinctions, and the in-sample-bank ablation provides relevant evidence. However, their relationship to this prior work should be explained to establish the methodological advance.

  2. Table 1 reports a final nMSE of 0.4658, whereas Figure 3(b) labels the Historical endpoint as 0.590. Appendices A.2 and A.4 distinguish Informs-Strict from an all-split STAP-Compatible reference without clearly identifying which results use each setting. Protocol, evaluation partition, and target normalization therefore need clarification before the reported 19.77% improvement can be assessed as a matched comparison.

  3. The paper validly defines empirical contributions as held-out prefix-risk differences and distinguishes them from causal effects or mutual information. Nevertheless, these gains can combine newly available evidence with corrections to earlier estimation errors. In Equation (4), G(C,X) can correct an underfitted context estimator using C alone, with that gain assigned to the Multimodal stage. Cross-fitting does not itself separate these sources, and Appendix C.5 does not examine this failure mode.

  4. The matched comparisons require more implementation and tuning detail. Appendix A.2 supplies several final configurations, and Appendix C.4 specifies control formulas. However, the one-shot estimator F(C,X,H), its representation of historical inputs, and the SMPD context estimator remain insufficiently described. Hyperparameter search ranges and operational definitions of equal capacity and tuning budgets are also needed to assess whether Table 2 isolates the proposed training principle.

Questions for authors
  1. How does SSPM extend ResMem-style residual correction? Please explain how the existing in-sample-bank control relates to this approach and which results establish the additional contribution.

  2. Which protocol and evaluation partition produced the Informs results in Table 1 and Figure 3(b)? Please specify normalization and historical-reference eligibility and reconcile 0.4658 with 0.590.

  3. How sensitive are the attribution shares to prefix-estimator fit? Can a context-only residual refiner recover part of the gain assigned to the Multimodal stage, thereby quantifying the contribution of earlier estimation errors?

  4. Please specify the one-shot estimator, its historical-input representation, the SMPD context estimator, and the hyperparameter search ranges. How were equal capacity and tuning budgets enforced?

3 comments
AuthorsOpenReview ↗

Author Response to xSFr (Part 1)

Thank you for your careful and constructive review. Your questions helped us sharpen the distinction from ResMem and clarify the evaluation protocols and implementation details.

Comment 1

[Q1] Novelty Beyond ResMem

Re: Residual retrieval itself is not new; SSPM’s contribution is to model where predictability emerges across information scales.

(1) Method. ResMem uses a base predictor plus residual memory. SSPM instead:

  • Builds Context, Multimodal, and Historical stages connected by OOF residual targets.
  • Makes each stage learn the preceding residual.
  • Measures held-out risk reductions across retained prefixes.

(2) Controlled evidence on MicroLens.

  • Residual-bank construction: We fix the base predictor and retrieval settings and change only the bank. Table 2’s in-sample-bank ablation tests this choice within SSPM; the new matched control isolates it in a fixed ResMem-style setup. The adapted ResMem-OOF and matched SSPM configurations are identical in this test.
  • Scale-wise learning without Historical memory: Both models use the same inputs and 500-tree budget. Scale-wise learning beats one-shot learning for all three paired seeds, showing a contribution independent of residual retrieval.

(3) Output. SSPM retains intermediate predictions and attributes held-out risk reductions to information scales (Figure 3), revealing where gains occur rather than reporting only an endpoint prediction.

We will cite ResMem and clarify these distinctions.

ControlComparison (MAE ↓)Improvement95% bootstrap CI
Residual bankIn-sample 1.1519 → OOF 1.14050.0114[0.0058, 0.0170]
Scale-wise learningOne-shot 1.1529 → Scale-wise 1.13990.0130[0.0007, 0.0249]

[Q2] Informs Protocol and Numerical Discrepancy

Re: Both values are correct but use different protocols:

  • Table 1 / STAP-Compatible: 0.4658.
  • Figure 3(b) / Strict Temporal V2: 0.5904.

Each test set has 258 samples, with only 28 overlapping; both results were reproduced from saved predictions. All 5,160 Strict Temporal V2 retrieval relations satisfy temporal eligibility. The reported 19.77% is a retrospective comparison with STAP’s published 0.5806 under STAP-Compatible, not a verified sample-matched improvement. We will label both protocols and disclose the normalization difference.

SettingSTAP-Compatible (Table 1)Strict Temporal V2 (Fig. 3b)
Split1,588 / 2581,270 / 318 / 258
NormalizationFull-data target statisticsTraining-only statistics
Historical referencesRetrospective referencesStrictly earlier samples
nMSE0.4658200.590373
AuthorsOpenReview ↗

Author Response to xSFr (Part 2)

Comment 2

[Q3] Attribution Sensitivity to Prefix Estimation

Re: Cross-fitting does not eliminate estimation error, and empirical attribution can depend on prefix-estimator quality. We added two MicroLens controls using the same information sources and paired validation samples; they do not replace the frozen SSPM model.

  1. Context-only residual repair: With a fixed Context predictor and identical OOF residual targets, we train capacity-matched Context-only and Context-plus-Multimodal refiners. Context-only repair recovers 8.93% of the Multimodal gain, and its bootstrap interval crosses zero. Multimodal refinement still gives ΔMSE = 0.1243 beyond repaired Context (95% CI [0.0804, 0.1687]).
  2. Prefix-estimator sensitivity: Stronger Context predictors reduce the absolute Multimodal gain, but its share remains 86.55%–89.77% across nine runs (three configurations × three seeds), and Multimodal is always dominant.

Thus, attribution magnitude depends on estimator fit, while the dominant-scale result is stable in the tested settings. We will describe contributions as model-dependent held-out risk reductions, not pure information gains or causal effects.

VariantnMSE ↓MSE reduction ↓
Context0.9095—
+ Context-only repair0.90450.0122
+ Multimodal refinement0.85340.1365

[Q4] Implementation Details and Fair Comparisons

Re: We will add the following implementation details and capacity-matched comparison:

  • Context: Historical popularity, calendar, and topic statistics, with optional user history; four Ridge/HistGradientBoosting configurations selected by validation MAE; eight chronological OOF blocks with expanding-window training.
  • Multimodal: 175-dimensional representation with a three-seed CatBoost ensemble.
  • Historical: Temporally eligible residuals.
  • Search: M1/M2 use K ∈ {3,5,10,20} and α ∈ {0.05,0.10,0.20,0.25,0.35}; selected M0 uses K = 5 and α = 0.25.
  • Fair comparison: The MicroLens control uses identical nonhistorical inputs, splits, seeds, and a total 500-ExtraTrees budget; neither method uses Historical retrieval. The result supports scale-wise learning beyond capacity differences.

These details, search settings, and controls will be added to the appendix.

The authors’ response has addressed my concern to some extent, but not sufficiently to change my overall assessment. I therefore maintain my original score.

Comments on the paper

3 comments
AuthorsOpenReview ↗

Author Response to ij33 (Part 1)

Thank you for your careful and constructive review. Your comments helped us clarify the theoretical scope, experimental controls, and presentation.

[W1] Theoretical Foundation and Cross-Fitting

Re: Proposition 1 supports scale-wise learning by decomposing Bayes-optimal prediction increments under nested information sets; it does not claim exact orthogonality of fitted estimators. SSPM implements this principle through:

  1. Restricted inputs: Context uses C, Multimodal adds X, and Historical accesses residual memory.
  2. OOF residual targets: Each stage learns the held-out errors of its preceding predictor.
  3. Frozen prefixes: Earlier predictors remain fixed during later refinement.

These mechanisms reduce in-sample fitting bias and enable structured residual learning, without requiring exact orthogonality. Stage contributions are measured by observed held-out risk reductions, so attribution remains directly measurable when fitted estimators differ from their Bayes-optimal counterparts. We will clarify this scope.

[W2] Historical Information vs. Estimation Improvement

Re: Historical refinement accesses label-informed residual memory beyond query-side C and X; unlike Multimodal refinement, it retrieves OOF residuals from historical instances. Our controls show:

  • A same-information refiner recovers only 19.8% of the Historical gain.
  • Historical retrieval gives the larger improvement (95% bootstrap CI [0.0061, 0.0182]).
  • Random neighbors and shuffled residuals degrade performance.

Thus, the tested gain comes from structured access to relevant historical residuals, not simply greater model capacity or arbitrary memory. We will frame it as an empirical gain from training-data memory—not new mutual information conditional on the full training data—and include training data and estimator randomness in the formal interpretation.

MicroLens validationMAE ↓MAE gain ↑
Global prefix1.152685—
Same-information refiner1.1502750.002410
Historical retrieval1.1405380.012147
Random neighbors1.157351−0.004666
Shuffled residuals1.156025−0.003341

[W3] Robustness of Scale Attribution

Re: SSPM attribution is held-out prefix-risk reduction, not intrinsic or causal information contribution; it depends on the information order and model configuration.

  • Context-only residual repair recovers 8.93% of the measured Multimodal gain.
  • Across nine Context-strength runs, the Multimodal share remains 86.55%–89.77% and is dominant in every run.

Absolute contributions vary with estimator quality, but the dominant-scale finding remains stable in the tested configurations. We will state this explicitly.

[W4] Evaluation Protocol Comparability

Re: The two Informs values come from different protocols, not inconsistent calculations:

  • STAP-Compatible: nMSE = 0.4658; retrospective evaluation with full-data target normalization.
  • Strict Temporal V2: nMSE = 0.5904; training-only normalization and temporally eligible references.

Both values were reproduced from saved predictions. Each evaluation set contains 258 samples, with only 28 shared; all 5,160 Strict Temporal V2 retrieval relations satisfy temporal eligibility. The reported 19.77% is a retrospective numerical comparison with STAP’s published result, not a verified sample-matched improvement. We will label both protocols and disclose the normalization settings and prior inspection of two test labels.

AuthorsOpenReview ↗

Author Response to ij33 (Part 2)

[W5] Label-Retrieval Degradation

Re: The label-retrieval control uses a convex mixture of prefix predictions and retrieved labels; under the tested configuration, it performs worse than residual refinement because the targets differ:

  • Label interpolation: combines full predictions with neighbor outcomes.
  • SSPM: retrieves errors remaining after Context and Multimodal refinement.

This supports the residual-target design, but the gap may also depend on interpolation weights and calibration. We will not attribute the entire degradation to double-counting.

[W6] Instance-Level Visualization

Re: Figure 4 shows the magnitude of stage-wise adjustments, not predictive correctness. We will label the absolute values as normalized correction magnitudes and add signed corrections.

[W7] Methodological Novelty

Re: SSPM shifts popularity prediction from endpoint optimization to scale-structured predictability analysis:

  • It organizes Context, Multimodal, and Historical evidence into ordered stages connected by OOF residual targets.
  • Retained prefixes directly measure held-out risk reductions and reveal where predictive gains emerge.
  • Unlike ICPF and STAP, the goal is not only retrieval-based accuracy but also interpretable scale-wise attribution.

A capacity-matched MicroLens control supports this principle: without Historical retrieval, scale-wise learning reduces MAE from 1.1529 to 1.1399 (95% CI [0.0007, 0.0249]). We will clarify these distinctions.

[W8] Presentation Clarity

Re: We will move key implementation details into the main text, reduce repeated terminology, and simplify the qualitative figure.

AuthorsOpenReview ↗

Author Summary for the Area Chair

We sincerely thank the Area Chair and all three reviewers (R1: ij33, R2: xSFr, R3: st2J) for their careful evaluation and constructive discussions. The rebuttal has helped us clarify SSPM's methodological contributions, strengthen its experimental evidence, and improve implementation transparency.

1. Core Contributions: Opening the Black Box of Popularity Prediction

SSPM opens the black box of popularity prediction. Rather than only improving endpoint accuracy, it measures predictive gains across Context, Multimodal, and Historical scales.

Its contributions are:

  • Scale-structured learning: OOF residual targets connect successive prediction stages.
  • Historical residual refinement: Historical retrieval corrects unresolved errors instead of directly retrieving full labels.
  • Predictability attribution: Retained prediction prefixes measure the held-out risk reduction associated with each information scale.

SSPM therefore goes beyond predicting popularity. It provides a structured way to examine where predictive gains emerge and how different information sources contribute to the final prediction.

2. Strengths Recognized by the Reviewers

We appreciate that all three reviewers recognized important aspects of our work.

  • R1 (ij33) recognized the statistical rationale behind OOF residual construction, the validity of the theoretical foundation, and the value of controlled experiments. The review also acknowledged the distinction between predictive risk attribution and causal attribution.
  • R2 (xSFr) highlighted the diagnostic value of retaining intermediate predictions, the sensible use of cross-fitting, and the evidence provided by matched controls and uncertainty analysis.
  • R3 (st2J) appreciated the information-scale perspective, the coherent progressive refinement design, and the insight provided by identifying where predictive gains emerge. The review also recognized the potential usefulness of this perspective beyond popularity prediction.

These assessments support our central motivation: making popularity prediction more transparent and analytically informative, rather than treating it solely as endpoint optimization.

3. Main Concerns and Additional Experimental Evidence

We conducted targeted experiments to address the main concerns raised during the discussion.

ConcernAdditional Evidence
R1/R2: Novelty beyond residual learning and ResMemOOF residual memory improves MAE from 1.1519 to 1.1405. Capacity-matched scale-wise learning improves MAE from 1.1529 to 1.1399 without Historical retrieval.
R1/R2: Reliability of attributionContext-only repair recovers 8.93% of the measured Multimodal gain. Across nine sensitivity runs, Multimodal remains dominant (86.55%–89.77%).
R3: Generality beyond tree-based learnersA trainable Transformer fusion head achieves lower MSE in all seven paired runs. Mean MSE decreases from 2.1896 to 2.1322 (2.62%).
R3: Feature construction and robustnessThe SMPD topic ablation improves validation MAE from 2.4123 to 2.3240. CLIP + VideoMAE also improves Historical retrieval performance on MicroLens.

These experiments strengthen the evidence that SSPM's benefits extend beyond a particular residual-memory design or tree-based backbone.

Importantly, we also clarified the interpretation of attribution. Theoretical orthogonality concerns Bayes-optimal predictors. Empirical contributions are directly measured held-out risk differences, conditional on the declared model and information configuration.

4. Reproducibility and Manuscript Revisions

Following R3 (st2J)'s detailed questions, we clarified the actual model families, encoder configurations, feature assignments, clustering pipeline, and historical-reference eligibility.

We also explained SSPM's complementary relationship to tree-based, Transformer-based, and graph-based methods, and added a controlled neural-backbone experiment.

The revised manuscript will make these implementation details and additional controls more accessible.

5. Concluding Remarks

We respectfully request the Area Chair to consider the following points when making their final recommendation: (i) the reviewers' recognition of SSPM's sound theoretical motivation, statistically grounded residual construction, and value in revealing where predictive gains emerge; (ii) the substantial new experimental evidence addressing concerns about methodological novelty, attribution reliability, and generality beyond tree-based learners; and (iii) our detailed clarifications of evaluation protocols, implementation choices, and reproducibility.

We believe these responses reinforce SSPM's central contribution: making predictive gains across information scales measurable and interpretable.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.