PaperlayerView exact revision on arXiv

Source-linked AI summary

Tabular Foundation Models for Multi-View Information Cascade Popularity Prediction

Wenting Zhu, Chenghua Gong, Sanchuan Guo, Chaozhuo Li, Yueyue Zhang, Xi Zhang

arXiv:2608.25048v1cs.SI

TL;DR

Existing popularity-prediction methods incompletely model heterogeneous views and high-order cross-view interactions. TFM4POP unifies static views with a tabular foundation model and models dynamics with a Neural-ODE branch, achieving consistent gains over baselines across datasets, observation settings, and metrics.

  • Problem

    Existing methods rarely model all four information views jointly and typically use shallow fusion, limiting their capture of high-order cross-view interactions.

  • Method

    TFM4POP converts non-tabular views into compact feature columns for unified TFM reasoning, separately encodes continuous-time dynamics with Neural ODEs, and fuses both representations via cross-attention.

  • Results

    TFM4POP consistently outperforms all baselines across both datasets, observation settings, and metrics, improving MSLE by 2.01%–12.76% and MAPE by 2.88%–18.12% over the best baselines.

  • Takeaways & Limitations

    The framework and benchmark support multi-view popularity prediction under more comprehensive, context-rich diffusion settings.

  • Takeaways & Limitations

    Existing benchmarks often lack textual and event-level context, limiting evaluation of richer diffusion settings.

Abstract

from arXiv · show

Predicting the future popularity of information cascades is essential for understanding information diffusion on social media. Despite recent advances, existing methods face two key limitations: they focus primarily on the cascade view while overlooking other information views that drive user engagement, such as textual semantics, visual content, and tabular attributes; and they fail to capture high-order cross-view interactions. To address these issues, we propose \textbf{TFM4POP}, the first framework to introduce tabular foundation models (TFMs) into popularity prediction, leveraging their pre-trained tabular priors to unify the modeling of multiple heterogeneous information views. Specifically, TFM4POP adopts a dual-branch design: the static branch employs a TFM as the feature-encoding backbone that jointly reasons over all static views through in-context learning to produce the static cascade representation, while the dynamic branch captures the continuous-time cascade dynamics with a dedicated Neural-ODE-based encoder. The two representations are then fused via cross-attention for the final prediction. Furthermore, to adapt the TFM to real cascade distributions, we apply parameter-efficient IA3 fine-tuning, achieving performance competitive with or better than full fine-tuning while updating substantially fewer parameters. In addition, we construct a comprehensive multi-view cascade benchmark that covers all four information views. Extensive experiments show that TFM4POP consistently outperforms state-of-the-art baselines across multiple datasets and observation settings.

1 Introduction

Information cascade popularity prediction supports applications such as misinformation control and trending-topic detection, but existing methods incompletely cover heterogeneous views and miss high-order cross-view interactions. TFM4POP addresses these limitations by unifying static views for TFM reasoning, modeling continuous-time dynamics separately, and adapting the TFM with parameter-efficient fine-tuning.

  • Background: Information cascade popularity prediction forecasts future cascade popularity from early-stage diffusion patterns and supports applications including misinformation control, trending-topic detection, and online marketing.
  • Limitations: Existing methods primarily focus on the cascade view, while independently encoding multiple views and combining them shallowly, which misses high-order cross-view interactions.
  • Challenges and solutions: TFM4POP addresses the modality gap by compressing non-tabular views into compact feature columns alongside raw tabular attributes, preventing feature-dimensionality imbalance from suppressing tabular signals.
  • Challenges and solutions: TFM4POP separately handles irregular continuous-time cascade diffusion with a dedicated dynamic branch, then integrates it with the static representation for popularity prediction.
  • TFM4POP: TFM4POP unifies four information views in a tabular input, allowing a pre-trained TFM to jointly reason over them through in-context learning.
  • Contributions: IA3 parameter-efficient fine-tuning adapts TFMs to real cascade distributions while preserving pretrained priors, achieving performance on par with or better than full fine-tuning.

2 Related Work

Existing information cascade popularity prediction methods include feature-based, statistics-based, and deep learning-based approaches, each with distinct limitations in feature dependence, accuracy, or representation learning.

  • Information Cascade Popularity Prediction: Feature-based methods [9] depend on hand-crafted feature quality and generalize poorly, while statistics-based methods [47] [41] offer interpretability but limited accuracy.Feature-based approaches rely on hand-crafted features; statistics-based approaches model diffusion with Poisson [47] or Hawkes [41] processes.
  • Information Cascade Popularity Prediction: Deep learning-based methods [5] [30] combine graph and sequence models to learn expressive cascade representations.

3 Preliminaries

The preliminaries formalize information cascades and their propagation graphs, then define four complementary views of each post. They motivate tabular foundation models, instantiated with TabPFN, for jointly modeling heterogeneous tabular signals.

  • Cascade Definitions: An information cascade C_i is an ordered sequence of user-timestamp tuples, beginning with the original publication and followed by chronological reposts.
  • Cascade Definitions: For a cascade observed until t_o, the cascade graph G_i(t_o) represents participating users as nodes and repost relationships as timestamped directed edges.
  • Information Views: Each post contains four complementary views: cascade structure, metadata and handcrafted cascade statistics, textual content T_i, and visual content I_i, which may be unavailable.
  • Tabular Foundation Models: Because existing methods partially or shallowly fuse these views, tabular foundation models provide a natural substrate for jointly modeling their heterogeneous signals in one table.
  • Tabular Foundation Models: TabPFN uses a table-specific transformer that embeds cells and applies row-wise inter-feature attention plus column-wise inter-sample attention while handling Boolean, categorical, numerical, and missing values.

4 Methodology

TFM4POP combines a unified static multi-view representation from tabular foundation models with a continuous-time dynamic representation from a Neural-ODE encoder. Cross-attention fuses the static profile with the dynamic trajectory, while IA3 and leakage-free context construction enable efficient, valid training.

  • Efficient and Leakage-Free Training: Training updates only IA3 vectors, the dynamic encoder, cross-attention, and predictor, while out-of-fold context construction prevents cascades from attending to their own labels.The TFM backbone remains frozen, and PCA is fitted on training data before application to all splits.
  • Static Multi-View Encoding: The static branch reduces modality gaps by compressing unstructured views, concatenating them with raw tabular features, and applying TabPFN attention across rows and columns.Row-wise attention captures within-cascade cross-view interactions, while column-wise attention relates cascades to similar historical examples.
  • Overall Architecture: TFM4POP uses dual branches: a TFM encodes textual, visual, structural, and tabular views, while a Neural-ODE encoder models continuous-time diffusion dynamics.The branches produce complementary static and dynamic cascade representations before fusion.
  • Dynamic Cascade Encoding: The dynamic branch represents repost events with structural and temporal features, combining continuous ODE evolution between irregular events with GRU jumps at event arrivals.Post-observation ODE propagation preserves elapsed silence, and the full latent trajectory retains stage-wise diffusion evolution.
  • Cross-View Fusion: Cross-attention lets the static representation query dynamic trajectory states, producing a profile-conditioned summary that is concatenated with the static token for MLP prediction.This avoids collapsing the dynamic representation into a single vector before fusion.

5 Experiments

Across Twitter and EventCas, TFM4POP consistently outperforms thirteen baselines under all observation settings and metrics, while ablations and sensitivity analyses show that multi-view inputs, the pretrained tabular prior, IA3 adaptation, and dynamic modeling materially affect performance.

  • Experimental setup: The experiments use Twitter and EventCas and compare TFM4POP against thirteen baselines spanning feature-based, statistics-based, deep-learning, and LLM-based methods.Twitter contains multimodal cascades with user profiles and textual and visual content, while EventCas is an event-centric multimodal benchmark curated from Sina Weibo.
  • Overall comparison: TFM4POP consistently outperforms all baselines on Twitter and EventCas across observation settings and metrics, improving MSLE by 2.01%–12.76% and MAPE by 2.88%–18.12% over the best baselines.The gains hold across all five random seeds and are most pronounced on EventCas.
  • Ablation study: Removing any single information view degrades performance, with MSLE increasing by up to 27.70%, confirming that the four views provide complementary signals.The ablations remove tabular, visual, structural, rationale, or dynamic inputs from TFM4POP.
  • Ablation study: Replacing the TFM backbone with a parameter-matched MLP causes the largest ablation degradation, increasing MSLE by 14.53% on Twitter and 70.76% on EventCas.This result attributes the improvement to the pretrained tabular prior rather than solely to assembled multi-view features.
  • Hyperparameter sensitivity: Performance follows a U-shaped trend with visual and textual PCA dimensions, as excessive compression loses information while excessive dimensions crowd out tabular attributes.EventCas is more sensitive than Twitter, supporting the attention-imbalance concern.
  • Hyperparameter sensitivity: Adapting the last four ICL blocks with IA3 performs best on both datasets; freezing the backbone underperforms, while adapting all blocks slightly erodes the pretrained prior.Performance remains stable for fold numbers K≥5 but degrades when K≤3 reduces available in-context examples.

6 Conclusion

TFM4POP introduces tabular foundation models into information cascade popularity prediction by unifying heterogeneous static views in a tabular space for joint in-context reasoning.

  • 6 Conclusion: TFM4POP is presented as the first framework to introduce tabular foundation models into information cascade popularity prediction, unifying textual, visual, structural, and raw tabular views for joint in-context reasoning.The framework compresses textual, visual, and structural information into feature columns and assembles them with raw tabular attributes in a unified tabular space.

A Dataset Details · A.1 Limitations of Existing Datasets.

Existing popularity-prediction benchmarks provide limited information for modern multi-view modeling because they largely contain only social-graph and basic cascade metadata, while lacking textual content and strong links to real-world events.

  • A.1 Limitations of Existing Datasets.: Weibo and Twitter contain social graphs and basic cascade metadata but omit textual data such as posts, reposts, comments, and user-generated context; random sampling within fixed time spans also weakens links to concrete real-world events.These limitations reduce the availability of event-related information for popularity prediction.

A.2 Data Construction and Statistics.

EventCas is constructed from Sina Weibo by curating 148 hot events across diverse topics. Posts are filtered by retweet activity, and their chronological retweets and repost texts are used to reconstruct cascade structure.

  • Data Construction and Statistics: EventCas comprises 148 hot events from Sina Weibo spanning entertainment, sports, social incidents, and public affairs.Events were collected over the past two years using event-specific hashtags or keywords.
  • Data Construction and Statistics: The dataset retains posts with more than five retweets and reconstructs each cascade from chronologically ordered retweets and parsed repost texts.For each retained post, user IDs and timestamps of all retweets are recorded.

A.3 Dataset Anonymization. · B Baseline Details

The dataset anonymization procedure protects privacy by replacing identifying information with consistent pseudonymous identifiers while preserving cascade diffusion structure and social connectivity.

  • A.3 Dataset Anonymization.: A.3 Dataset Anonymization.: User IDs, usernames, post IDs, and mentioned usernames are replaced with consistently mapped pseudonymous identifiers, concealing personally identifiable information while preserving cascade traces and repost paths.The procedure maintains the integrity of diffusion structure and social connectivity.

B.1 Feature-based methods … C Tabular Feature Definitions

The paper reviews feature-based, statistical, and deep learning approaches to cascade popularity prediction, then defines tabular features computed from early cascade dynamics across four semantic aspects. These features use only information available within the observation window, with dataset-specific additions for Weibo.

  • B.1 Feature-based methods: Feature-P uses observed cascade size at t_o to predict future popularity P(t_p) from P(t_o).It also extracts handcrafted structural and temporal cascade features for Feature-Linear and Feature-MLP predictors.
  • B.2 Statistical-based methods: DeepHawkes represents cascades as diffusion paths and uses recurrent networks to model reposting self-excitation, temporal decay, and user influence.The learned dynamics support prediction of future cascade growth.
  • B.3 Deep learning-based methods: CasCN encodes time-ordered subcascade graphs with graph convolutions and uses an LSTM to model structural evolution for popularity prediction.Its representation combines local snapshot structure with temporal evolution.
  • B.3 Deep learning-based methods: CasFlow jointly captures local and global cascade structure while modeling propagation uncertainty to improve cascade growth prediction.MINDS instead uses sequential hypergraphs to model macroscopic and microscopic diffusion, cross-cascade interactions, and shared multi-scale representations.
  • B.3 Deep learning-based methods: CTCP [30] formulates cascade popularity prediction in continuous time and combines ODEs with observed cascade representations to guide diffusion-based generation of future popularity trends.The method uses continuous-time dynamics as conditions for prediction.
  • B.3 Deep learning-based methods: MMCas jointly encodes cascade dynamics, user profiles, textual content, and visual content, then fuses them with an interpretable reweighted mixture-of-experts mechanism.The multimodal design targets popularity prediction by modeling interactions among heterogeneous information sources.
  • C Tabular Feature Definitions: Table 5 lists 21 tabular features shared by Weibo and Twitter, extracted solely from early cascade dynamics observed within t_o and organized into four semantic aspects.Weibo additionally includes poster/resharer historical activity and social-neighbourhood features unavailable in Twitter.
Loading 2608.25048v1…